SPIGEN: A Configurable Population-Coded Spike Pattern Generator for Single-Spike Learning in Spiking Neural Networks

SPIGEN: A Configurable Population-Coded Spike Pattern Generator for Single-Spike Learning in Spiking Neural Networks | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article SPIGEN: A Configurable Population-Coded Spike Pattern Generator for Single-Spike Learning in Spiking Neural Networks Zübeyir Özcan, İlknur Kayıkçıoğlu Bozkır, Temel Kayıkçıoğlu This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6709572/v1 This work is licensed under a CC BY 4.0 License Status: Under Revision Version 1 posted 5 You are reading this latest preprint version Abstract This study introduces SPIGEN, a configurable spike pattern generator designed to convert continuous-valued input features into fixed-length binary spike representations suitable for spiking neural networks. Based on a population coding framework, SPIGEN enables flexible configuration through multiple quantization and encoding schemes. Its primary goal is to transform continuous features into time-independent spike patterns that preserve inter-class separability and intra-class consistency. The study includes two experimental stages to evaluate the effectiveness of the proposed method. First, SPIGEN was used to generate spike patterns from two benchmark datasets—IRIS and Digits—and a synthetically generated Gaussian dataset. The resulting spike patterns were classified using conventional machine learning algorithms and a single-layer perceptron to examine how well the transformation process retained the information content of the original features. The results indicated that SPIGEN patterns yielded classification performances comparable to the original features, and the choice of encoding configuration significantly influenced the outcome. In the second stage, the generated spike patterns were applied as synaptic inputs to biologically inspired neuron models, where '1' indicates a spike and '0' denotes silence. When both excitatory and inhibitory synapses were used during training, the evaluation results suggested that SPIGEN is a viable solution for representing continuous input features in biologically plausible spiking neural networks. Spike pattern generation spiking neural networks population coding feature encoding supervised learning biological neuron models neuromorphic classification Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 1. Introduction The information processing capabilities of the brain are primarily realized through a vast network of interconnected neurons that communicate using spikes. Each neuron receives input in the form of structured spatiotemporal patterns transmitted via synaptic connections, processes this information according to its internal dynamics, and relays output to downstream neurons [ 1 ]. Due to its computational speed, sparsity, and energy efficiency, this mechanism has drawn significant attention from researchers in neuroscience, artificial intelligence, and computer science [ 2 , 3 ]. Attempts to emulate this form of computation have led to the development of biologically inspired models known as Spiking Neural Networks (SNNs). Over the past two decades, SNNs have found increasing applications in various domains, including neuromorphic hardware, robotics, and real-time sensory data processing [ 4 ]. Although spiking neurons operate with spikes, real-world inputs are continuous. Biological systems resolve this via sensory organs, while artificial systems requires spike encoding to fill in this gap. As a result, various biologically inspired methods have been developed to convert continuous data into spike-based representations [ 2 ]. Broadly, these methods fall into two main categories: rate-based and temporal encoding strategies. Rate coding conveys information through the number of spikes emitted over a fixed time window, and it includes variants such as count rate coding and density rate coding. In contrast, temporal coding relies on the precise timing of spikes to represent stimulus characteristics. Several subclasses of temporal coding have been proposed in the literature, including temporal contrast [ 5 ], correlation [ 6 ], deconvolution [ 7 , 8 ], latency/ISI [ 9 ] and global reference [ 10 , 11 , 12 ] based approaches. These diverse encoding paradigms have been applied across a range of SNN tasks, including sensory signal representation, temporal pattern recognition, and real-time event-driven neuromorphic computation [ 3 ]. In both rate and temporal coding strategies, continuous-valued stimuli are converted into spike trains by individual neurons, each encoding specific features of the input. In contrast, a third approach known as population coding represents stimuli through the collective activity of groups of neurons [ 13 ]. In this approach, real-world stimuli are transformed into meaningful spatiotemporal spike patterns by multiple neurons whose dynamics are selectively tuned to different aspects of the input. Unlike previously mentioned encoding schemes, population coding distributes information across multiple synaptic channels which allows a richer and more robust representation [ 14 ]. Recent studies have provided substantial evidence that information is represented through the coordinated activity of neuronal populations across various brain regions. Panzeri et al. [ 15 ] showed that population coding happens at different levels in the brain. In the auditory cortex, Pachitariu et al. [ 16 ] showed that how neuron populations respond to sounds depends on the brain’s state. Stringer et al. [ 17 ] studied how groups of neurons respond to visual input in the mouse brain. Panzeri et al. [ 18 ] studied how noise affects groups of neurons. They found that the amount of noise has an important impact on how well the brain can use population coding to carry information. In addition to spike encoding, the learning phase determines how synaptic weights are adjusted and governs the functional behavior of SNNs in performing various tasks. A wide range of learning approaches has been proposed for training SNNs. These methods can be divided into different subgroups based on network type, how learning is done, how information is encoded, how the structure changes, and how knowledge is represented [ 19 ]. One of the earliest supervised learning methods, SpikeProp, adapts the classical backpropagation algorithm to spiking neurons with single spike outputs [ 20 ], leading to multiple variants like QuickProp and Multi-SpikeProp [ 21 ]. ReSuMe, combines STDP and anti-STDP to adjust synaptic weights based on mismatch between desired and actual spike trains [ 22 ]. The Tempotron algorithm updates weights based on postsynaptic potential thresholds by capturing essential aspects of spike timing [ 23 ]. Reward-modulated STDP introduces global modulatory signals, mimicking dopamine-like mechanisms to guide synaptic plasticity based on delayed rewards [ 24 ]. SuperSpike allows gradient-based training in multilayer SNNs by approximating the non-differentiable spike function with smooth surrogate functions [ 25 ]. In single-spike learning paradigms, where each neuron receives at most one spike per synapse, spike pattern generators are essential in encoding continuous inputs into sparse but informative spatiotemporal spike patterns. Several recent studies have combined population coding with single-spike learning to address this challenge. Kheradpisheh et al. [ 26 ] proposed a rank-order coding strategy in which spike order encodes the input. Similarly, Liu et al. [ 27 ] employed a first-spike decision model, and Tang et al. [ 28 ] applied structured population codes in reinforcement learning. Moldwin and Segev [ 29 ] demonstrated that binary input patterns can be successfully learned and classified by a biologically detailed pyramidal neuron. They adopted a biologically motivated constraint suggested by Amit et al. [ 30 ], in which all synaptic weights are constrained to be positive-valued. As the input patterns were inherently binary, the system did not require a separate spike pattern generator. Their results showed that, despite the complexity and biophysical constraints of real neurons, such binary-coded inputs could be effectively separated through synaptic weight adaptation. A subsequent study by Kayıkçıoğlu Bozkır et al. [ 31 ] adopted the same perceptron-based learning mechanism and investigated the classification performance of a pyramidal neuron model on real-world ECG data. They used Gray coding to convert analog ECG signals into binary spike patterns and assumed that all synaptic connections were excitatory. Their proposed mirroring approach led to a significant improvement in classification accuracy. Nevertheless, they did not investigate the potential impact of using alternative input encoding strategies in place of Gray coding on classification outcomes. This study proposes SPIGEN, a spike pattern generator, which addresses the challenge of encoding continuous-valued features for use in spiking neural networks. Unlike conventional rate or temporal encoding methods, the proposed approach transforms analog features into fixed-length binary spike patterns using a configurable pipeline consisting of quantization, coding scheme, and bit-length parameters. The main contributions of this study can be summarized as follows: It introduces a flexible spike encoding architecture based on population coding, which supports multiple quantization and encoding schemes. It provides a comprehensive evaluation of three quantization methods combined with three encoding strategies and bit length. It analyzes how spike-based representations interact with both conventional classifiers and biologically inspired neuron models It incorporates both excitatory and inhibitory synaptic connections in the learning process, enhancing the biological plausibility. 2. Materials and Method 2.1. Datasets Three different datasets were used in this study. Each dataset had a specific purpose. The IRIS and Digits datasets were used as standard benchmarks [ 32 , 33 ]. These datasets helped evaluate classification performance under different encoding strategies. Additionally, a synthetic Gaussian dataset was used for visualization. It allowed the examination of spatial structure and class separability in two dimensions. As the first dataset, the IRIS dataset includes 150 samples where each has four continuous features. The samples are grouped into three classes based on iris species [ 32 ]. Secondly, The Digits dataset was taken from the scikit-learn library. It includes 1797 grayscale images of handwritten digits from 0 to 9. Each image is converted into a 64-dimensional feature vector [ 33 ]. Lastly, a synthetic dataset was created using the make_blobs function from scikit-learn [ 34 ]. It contains 1600 samples in two dimensions. The samples are grouped into eight Gaussian clusters. Each cluster has some overlap with others, controlled by a standard deviation of 0.6. 2.2. Methods This study employes a computational method for classifying continuous-valued feature vectors using a spiking neural network. The method includes spike-based encoding, neural modeling, and supervised learning in a single system. Figure 1 shows an overview of the full process. In the feature-to-spike pattern transformation phase, continuous-valued feature vectors are converted into spike patterns that spiking neural networks can process. This is done using SPIGEN, a spike pattern generator developed specifically for this study. The generator is trained using the statistical properties of the training data. After training, the same generator is used for both training and test data. In the network design phase, a spiking neural network is built to respond to spike patterns with class-specific outputs. The network has one layer of output neurons, and each neuron represents a class. Each bit in the spike pattern connects to all output neurons through two synapses: one excitatory and one inhibitory. The spike pattern acts directly as input to the output neurons through these synaptic connections. In the training and testing phase, the network is trained and tested using spike patterns created by SPIGEN. A supervised learning rule updates the synaptic weights during training. The synaptic weights are fixed once the training process is completed. In the testing stage, new spike patterns are given to the network using the same simulation settings. 2.2.1. SPIGEN This study proposes a method to generate spike patterns using population coding. The method converts continuous-valued feature vectors into spike patterns that can be interpreted by SNNs. It creates bit-wise spike patterns with fixed length. Each bit shows whether a spike is present, without depending on time. Figure 2 shows an example from the spike pattern generation process. The transformation begins with quantization. In this step, continuous-valued feature vectors are turned into a limited number of levels. The number of levels is set by the chosen bit length. To handle different data distributions, three quantization methods are used. Uniform quantization divides the feature range into equal intervals. Quantile-based quantization creates intervals with equal numbers of samples. K-means-based quantization groups values into clusters and assigns each to a level. After quantization, each level is turned into a spike code. Three encoding methods are used. In standard binary encoding, the level is written in standard binary form. In Gray encoding, the level is written so that only one bit changes between two nearby levels. In intensity encoding, a number of bits at the beginning are set to one. The rest of the bits are zero. This way, the number of spikes corresponds to the level of the quantized feature. The spike codes from all features are flattened to create a single spike pattern for each sample. These patterns are simple in structure but still carry the important information from the original features. 2.2.2. Classfiers and Spiking Neural Network In the study, two types of classification approaches were used to evaluate the performance of SPIGEN. In the first group, conventional machine learning algorithms were applied to both feature vectors and their corresponding spike patterns. These classifiers included k-Nearest Neigbors (k-NN) [ 40 ], Support Vector Machine (SVM) [ 41 ], Naive Bayes (NB) [ 42 ], Linear Discriminant Analysis (LDA) [ 43 ], and Single Layer Perceptron (SLP) [ 44 ], all implemented using the scikit-learn library with default settings. In the second group, SNNs were used to test how well the spike patterns support classification. The networks had only one output layer for each. Each neuron in the output layer represented a class. Each input sample is given to the network as a spike pattern. Each bit in the pattern acts as a presynaptic spike. These spikes are sent through fixed synaptic connections to all output neurons at the same time. To see how neuronal dynamics affect spike-based pattern processing, three neuron models were tested. Each model was used in a separate version of the same network. The input encoding, synaptic setup, and training method stayed the same for all models. 2.2.3. Neuron Models and Synaptic Connections When building SNNs, three biologically inspired spiking neuron models which are Hodgkin-Huxley (HH) [ 35 ], Leaky Integrate-and-Fire (LIF) [ 36 ], and Izhikevich (IZH) [ 37 ] were utilized. Each model was tested separately within the same network structure. The input encoding, synaptic setup, and learning procedure were kept the same across all cases. This allowed a fair comparison of how each neuron’s internal behavior affects classification. Table 1 Model parameters used in simulations. The table lists the biophysical parameters for the LIF, IZH, and HH neuron models, as well as the synaptic properties of AMPA and GABA synapses. Model Parameter Value Model Parameter Value LIF V rest -60 mV IZH V rest -60 mV V threshold -45 mV V threshold -40 mV V reset -60 mV V peak 35 mV V peak 35 mV C m 1 mF C m 0.45 mF k 0.7 nS/mV g 0.001 µS a 0.03 1/ms HH E Na 55.17 mV b -2 nS E K -72.14 mV c -50 mV E leak -49.42 mV d 100 pA g Na 1.2 µS AMPA E AMPA 0 mV g K 0.36 µS τ AMPA 5 ms g leak 0.003 µS GABA E GABA -70 mV C m 0.01 mF τ GABA 6 ms In this study, AMPA synapses were used to model excitatory connections [ 38 ], while GABA A synapses were used to represent inhibitory ones [ 39 ]. These synaptic types were chosen for their biological relevance and fast response characteristics. The specific parameter settings used for each model are summarized in Table 1 . Synaptic weights were initially set to zero, assuming no functional connection at the beginning of training. These values were then updated during an adapted perceptron rule. 2.2.4. Learning Rule In this study, an adapted version of the classical perceptron learning rule is employed to update synaptic weights [ 44 ]. In classical rule, the weight update is performed as follows: $$\:{{w}_{i}}^{New}\leftarrow\:{{w}_{i}}^{Old}+\varDelta\:{w}_{i}$$ 1 $$\:\varDelta\:{w}_{i}=\eta\:\times\:\left(d-out\right)\times\:{x}_{i}$$ 2 Here, \(\:out\) refers to the binary output of the neuron, \(\:d\) is the desired output, \(\:{x}_{i}\) is the input at position \(\:i\) , and \(\:\varDelta\:{w}_{i}\) denotes the change in the synaptic weight. According to this rule, the weights are updated proportionally to the classification error and the input activation. While this rule is applicable to artificial neurons without constraints on weight values, it requires modification for use with spiking neurons, especially in biologically inspired settings. Moldwin and Segev [ 29 ] highlighted that synaptic weights in biological neurons must remain non-negative, since negative values would imply non-physical conductance values. To enforce this constraint, they proposed a modified update rule. This rule ensures that if a weight becomes negative during training, it is instead clipped to zero. However, this formulation only applies to excitatory synapses and does not address inhibitory ones. Since the present study includes both excitatory and inhibitory synapses, Eq. 2 is extended to account for synaptic polarity and weight constraint. The generalized update rule is defined as follows: $$\:{{w}_{i}}^{New}\leftarrow\:\left\{\begin{array}{cc}max\left(0,\left({{w}_{i}}^{Old}+\varDelta\:{w}_{i}\right)\right),&\:For\:excitatory\:synapse\\\:max\left(0,\left({{w}_{i}}^{Old}-\varDelta\:{w}_{i}\right)\right),&\:For\:inhibitory\:synapse\end{array}\right.$$ 3 Equation 3 allows the learning rule to update weights while preserving the non-negativity and polarity constraints of both synapse types. Additionally, for constant input terms such as \(\:{I}_{base}\) ​, weight updates are applied according to Eq. 2 , as in standard perceptron learning. 2.2.5. Training and Testing The training process begins by selecting a training sample and converting it into a spike pattern using the previously constructed spike pattern generator. Each input in the pattern is associated with two predefined synaptic connections which are excitatory and inhibitory. The spiking neural network is then simulated for 50 ms. At the 10th ms of the simulation, input spikes are applied to synaptic connections corresponding to the inputs with a value of 1 in the spike pattern. During the simulation, the membrane potentials of all output neurons are recorded. An output neuron is labeled as active if it produces at least one spike, and inactive otherwise. If an output neuron fails to spike when it should, or spikes when it should not, its synaptic weights are updated using the learning rule. After processing a training sample, the procedure continues with the next one until all samples are used. One complete pass through all training samples is considered an epoch. If the network still produces incorrect outputs, training continues for additional epochs until the desired accuracy is achieved or a maximum epoch limit is reached. At the end of training, the weights are fixed and the system proceeds to the testing phase. Testing also consists of two stage which are simulation and decision. For each test sample, a spike pattern is generated using the same spike pattern generator used in training. The pattern is mapped onto the network using the same configuration and synaptic connections. Class labels are decided based on the outputs of the neurons that detect specific classes. For each test spike pattern, three different cases may happen. In the first case, only one output neuron produces spikes where the other neurons stay silent. In this case, the class of the active neuron is chosen as the label. In the second case, more than one output neuron produces spikes. Then, the label is taken from the neuron that spikes first. In the third case, no output neuron spikes. In this situation, membrane potentials are checked. The label is given based on the neuron that has the highest membrane potential. 2.3. Experimental Design and Simulation Environment The experimental procedure was carried out in two main stages. In the first stage, classical machine learning algorithms were applied directly to the raw analog feature vectors to evaluate their baseline classification performance. In the case of synthetic dataset, decision boundaries were visualized in two-dimensional space. For IRIS and Digits datasets, performance was evaluated using the F1-score. No spiking neuron models were used in this first stage. In the second stage, all datasets were passed through the spike pattern generation process. The resulting spike patterns were evaluated using the same classification methods as in the first stage. In addition, three spiking neuron models were included as SNN classifiers. For all spiking models, the number of training epochs was set to 500. For the spike pattern generator, bit length was selected as 5, quantization was selected as uniform, and encoding was selected as intensity coding by default. To assess the effect of encoding parameters on classification performance, the influence of bit length, quantization strategy, and encoding scheme was evaluated independently. In the case of the Digits dataset, which is already quantized into 16 levels, experiments related to quantization and bit length were excluded. For the IRIS and Digits datasets, the original data were shuffled and then split into two parts. 70% of the data were used for training, and 30% were used for testing. This process was repeated 5 times to get more reliable results. Each time, the data were shuffled and split again. The results from each shuffle were calculated separately. In the end, the average of all results was taken for evaluation. All simulations involving spiking neuron models were conducted using the NEURON simulation environment, which provides biophysically realistic modeling capabilities for neuronal dynamics [ 45 ]. The simulator was integrated with Python for data preprocessing, spike pattern generation, parameter configuration, and result analysis. The experiments involving conventional machine learning algorithms were conducted using Python’s scikit-learn library [ 34 ]. The model specific parameters were chosen as the default configurations defined by the library. 3. Results 3.1. IRIS Dataset The IRIS dataset was used to test how spike pattern generator configurations affect classification performance. Different types of classifiers were included in two separate experiments. In the first experiment, all encoding schemes and quantization methods were combined and tested while the bit length was kept as 5. In the second experiment, bit length was iterated from 2 to 7, while uniform quantization and intensity encoding were used. In both experiments, the F1-score was used to measure performance. Results were calculated for both classical classifiers and spiking neural network models and mapped on Table 2 and Table 3 . Table 2 shows the performance of classical machine learning models. Continuous features gave the highest f1-scores for all classifiers except for SLP. These scores were used as the performance benchmark. While SVM had the best score with 98.2%, LDA and k-NN classifiers followed with 99.1%, and 96.0%, respectively. Intensity encoding gave results close to the continuous baseline. This was clearer when it was used with uniform or quantile quantization. SLP showed high performance with intensity and Gray encoding, reaching up to 96.0%. NB has shown high sensitivity to encoding types with respect to baseline, especially when using binary encoding. With uniform quantization, its score even dropped to 80.5%. Results from SNNs have shown that intensity encoding promises the most stable and high results. When used with uniform or quantile quantization, HH, IZH, and LIF had f1-scores over 94%. These results were close to the scores of classical classifiers. Gray encoding showed more variation. Its performance changed with different quantization methods. The scores were between 85.6% and 92.4%. Binary encoding gave the lowest scores for SNNs. In the HH model, its score dropped to 63.1% with k-means quantization. Table 2 F1-score results (%) of IRIS dataset obtained using different combinations of encoding and quantization strategies across various classifiers. The “Continuous” row shows baseline results without encoding. Here the bit length was chosen as 5. Each value represents the average accuracy over five runs. Coding Quantization KNN SVM SLP NB LDA HH IZH LIF Continuous Continuous 96,0 98,2 92,0 94,7 99,1 Intensity Uniform 95,1 94,7 95,1 94,7 87,9 95,1 94,7 94,7 Quantile 95,6 93,8 96,0 93,3 89,9 95,1 93,8 94,7 k-Means 95,6 94,2 95,1 93,4 82,9 94,7 93,8 94,7 Gray Uniform 94,3 93,3 96,0 92,5 94,3 85,6 92,4 93,8 Quantile 91,1 94,2 90,2 90,8 92,1 74,0 87,6 90,0 k-Means 92,9 92,4 90,7 87,8 93,4 77,5 89,2 90,2 Binary Uniform 84,1 92,4 74,9 80,5 95,1 72,5 83,1 83,6 Quantile 85,0 93,8 87,0 85,3 95,6 66,4 87,4 88,4 k-Means 81,5 92,9 80,2 83,4 95,6 63,1 84,4 85,3 Table 3 F1-score results (%) of IRIS dataset obtained for different bit lengths used in the spike pattern generator. The “Continuous” row shows baseline results without encoding. Here, the quantization strategy and coding scheme were chosen as uniform and intensity, respectively. Bit Length KNN SVM SLP NB LDA HH IZH LIF Continuous 96,0 98,2 92,0 94,7 99,1 2 92,1 95,1 94,3 87,0 90,3 88,2 91,3 91,3 3 94,2 94,2 94,7 94,7 94,2 89,8 93,4 94,2 4 95,1 95,6 95,6 93,8 94,6 93,3 93,8 94,2 5 95,1 94,7 95,1 94,7 87,9 95,1 94,7 94,7 6 95,1 93,8 95,1 92,9 74,9 94,7 93,8 93,8 7 95,1 93,8 95,1 93,4 74,4 94,7 93,8 94,2 Table 2 shows that the encoding scheme and quantization method have a significant effect on classification performance. Intensity encoding gave better results than Gray and binary coding in most classifiers. Uniform and quantile quantization approaches gave higher f1-scores than k-means quantization. This was clearer when they were matched with intensity encoding. Binary encoding is simple, but it gave the lowest performance. This was more obvious when it was used with k-means quantization. Using intensity encoding with uniform or quantile quantization gives a strong combination for both traditional machine learning algorithms and spiking neural networks. Table 3 shows the f1-score results for different bit lengths in the spike pattern generator. As expected, increasing the bit length improved classification performance in most classifiers. But this improvement continued only up to a certain point. There was a clear increase from 2-bit to 4-bit lengths. This was more visible in SVM, NB, and the spiking neuron models. For example, the HH model increased from 88.2% at 2 bits to 93.3% at 4 bits. However, after 5 bits, performance stopped improving. In some cases, like LDA and NB, it even started to drop. According to table, choosing the bit length as 4 or 5 seems to give the best balance between computational load and classification performance. 3.2. Digits Dataset The Digits dataset was used to evaluate the impact of different encoding schemes on classification performance. Since the input feature vectors in this dataset are already quantized into 16 discrete levels, further quantization or bit length variation was not necessary. Therefore, the experiments focused on comparing the performance of different encoding schemes. Classification was performed using both traditional machine learning models and spiking neural networks models, and the F1-score was used as the evaluation metric. Table 4 F1-score results (%) of Digits dataset for different encoding schemes applied to the same feature vectors. The “Continuous” row shows baseline results without encoding. Coding KNN SVM SLP NB LDA HH IZH LIF Continuous 98,0 97,7 95,3 83,4 94,9 Intensity 97,6 98,0 97,0 80,2 87,7 95,6 95,4 95,6 Gray 88,7 94,3 91,5 74,1 92,1 91,2 91,2 91,4 Binary 92,8 96,2 94,4 72,9 93,9 93,0 93,2 93,2 Table 4 presents the f1-score results for the Digits dataset obtained using different encoding schemes. The continuous feature representation achieved the highest overall performance in most classical classifiers, particularly with SVM and k-NN. Among the encoded variants, intensity encoding consistently outperformed Gray and binary across both traditional and SNNs. With intensity encoding, SLP reached 97.0%, overcoming its baseline. The HH, IZH, and LIF models also achieved scores above 95%. Gray and binary encoding resulted in lower performance. These results confirm that intensity encoding is the most effective scheme for preserving the discriminative structure of the input space when working with pre-quantized features such as those in the Digits dataset. 3.3. Synthetic Dataset The synthetic dataset was used to visually analyze how different encoding schemes and parameter configurations affect class separability in both traditional and spiking neural networks. Its two-dimensional structure and clearly clustered class distribution make it suitable for illustrating decision boundaries. Three figures are presented in this section. The first shows the decision regions obtained using SLP classifier for all combinations of encoding schemes and quantization strategies. Although only SLP results are visualized, results for additional machine learning classifiers were also evaluated and are included in the supplementary material. The second figure illustrates the same encoding and quantization configurations applied to the IZH model. Similarly, results for other spiking models were also generated and are provided in the supplementary material. The final figure examines the effect of varying bit length on classification boundaries across neuron models using fixed intensity encoding and uniform quantization. Figure 4 presents the decision boundaries obtained using the SLP for different combinations of encoding schemes and quantization methods. As expected, the continuous input yielded the cleanest and most compact class regions. Among the encoded inputs, intensity encoding produced the most coherent and stable decision boundaries, especially when paired with quantile or uniform quantization. Gray encoding resulted in moderately structured regions but exhibited irregular edges and fragmented zones. Binary encoding produced the noisiest and most scattered boundaries across all quantization methods. These results demonstrate that intensity encoding allows SLP to approximate continuous classification behavior more closely than other encoding strategies. Figure 5 shows the decision boundaries obtained from the IZH model results. Among the three encoding types, intensity encoding once again demonstrates superior performance, producing smooth and well-structured decision regions. Gray encoding provides moderate separability but suffers from fragmented and fuzzy boundaries. Binary encoding shows the weakest boundary structure, with highly scattered and disorganized regions. Compared to the SLP results, the decision surfaces obtained from the IZH model are generally coarser and less defined. Nevertheless, the IZH model, when paired with intensity encoding and proper quantization, is capable of capturing meaningful spatial structure. These results confirm the compatibility of the proposed spike encoding method with spiking neural networks. Figure 6 illustrates the effect of varying bit length on the decision boundaries generated by four different models. All configurations use intensity encoding and uniform quantization. Across all models, increasing the bit length generally improves the structure and clarity of the decision regions. With n = 2, decision boundaries are noisy, where class regions often overlap. As the bit length increases to 4 or 5, boundaries become sharper and more aligned with cluster distributions. For most models, n = 5 provides a good trade-off between resolution and generalization. Beyond 5 bits, the improvements tend to stop. These visualizations reinforce the findings from the IRIS dataset, showing that moderate bit lengths are sufficient to encode meaningful spike patterns across both conventional and SNNs. 4. Discussion 4.1. Effect of Quantization In the experiments, quantization strategy was found to have a significant impact on classification performance. Across both traditional classifiers and spiking neural networks, quantile and uniform quantization generally outperformed k-means clustering. This trend was evident in the IRIS dataset results, where intensity encoding paired with quantile or uniform quantization consistently achieved higher f1-scores across most models. Visual inspection of decision boundaries in the synthetic dataset further supports this finding. Class regions under quantile and uniform quantization appeared more structured and aligned with true class distributions. In contrast, k-means quantization led to irregular and fragmented boundaries. These results indicate that quantile and uniform quantization methods provide more stable and generalizable input representations. 4.2. Effect of Coding Scheme The choice of coding scheme had a pronounced effect on classification performance, both in terms of quantitative performance and decision boundary structure. Among the evaluated schemes, intensity encoding consistently delivered the highest f1-scores across datasets and classifier types. As shown in Table 2 and Table 4 , intensity encoding performed comparably to continuous representations. Visualizations from the synthetic dataset reinforced this trend. Intensity-encoded patterns produced well-separated class regions. Gray encoding demonstrated moderate performance, providing smoother transitions than binary but still suffering from instability under certain quantization methods. Binary encoding yielded the lowest performance overall, characterized by noisy and fragmented class boundaries. These results suggest that intensity encoding is more effective in preserving the structure of features. This makes it a robust scheme for spike patter generation. 4.3. Effect of Bit Length Bit length, which determines the resolution of the spike pattern representation, was shown to influence classification performance non-linearly. Results from the IRIS dataset and the synthetic dataset revealed that increasing the bit length from 2 to 4 or 5 generally led to significant improvements in both f1-score and decision boundary clarity. This trend was observed across both traditional classifiers and SNNs. Notably, performance gains plateaued beyond 5 bits, and in some cases, further increases led to declines. These findings indicate that moderate bit lengths are sufficient for capturing class-discriminative information in spike pattern generation, while excessively high resolutions may introduce unnecessary complexity without tangible benefit. 5. Conclusion This study introduced SPIGEN, a spike pattern generation framework, designed to convert continuous-valued feature vectors into fixed-length, biologically interpretable spike patterns. The method enables compatibility with both traditional machine learning classifiers and SNNs by incorporating configurable quantization, encoding, and bit-length parameters. Experimental results across multiple datasets showed that intensity encoding combined with uniform or quantile quantization offers the most robust classification performance, closely approximating continuous feature baselines. Moderate bit lengths, particularly 4 and 5 bits, were found to provide an optimal balance between resolution and performance. Visual analyses on a synthetic dataset further confirmed that the proposed encoding strategy preserves spatial class structure. These findings highlight the effectiveness of SPIGEN and its suitability for neuromorphic computing, embedded learning systems, and low-power spike-based processing. Future work may explore adaptive encoding mechanisms, online learning extensions, and hardware implementations to further enhance practical applicability. Declarations Funding: This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. Author Contribution Zübeyir Özcan: Conceptualization, Methodology, Software, Formal analysis, Investigation, Data curation, Visualization, Writing – Original Draft.İlknur Kayıkçıoğlu Bozkır: Validation, Writing – Review & Editing, Resources.Temel Kayıkçıoğlu: Supervision, Project administration, Writing – Review & Editing.All authors have read and approved the final manuscript. Data Availability: No new data sets were acquired in this study. Code and trained models will be made available on reasonable request. References Gerstner, W., Kistler, W.M., Naud ve, R.: L. Paninski, Neuronal Dynamics: From Single Neurons to Networks and Models of Cognition. Cambridge University Press (2014) Auge, D., Hille, J., Mueller ve, E., Knoll, A.: «A survey of encoding techniques for signal processing in spiking neural networks». Neural Process. Lett. cilt. 53 , 4693–4710 (2021) Forno, E., Fra, V., Pignari, R., Macii ve, E., Urgese, G.: «Spike encoding techniques for IoT time-varying signals benchmarked on a neuromorphic classification task». Frontiers Neurosci. cilt. 16 , 999029 (2022) Yamazaki, K., Vo-Ho, V.-K.: D. Bulsara ve N. Le, «Spiking Neural Networks and Their Applications: A Review,» Brain Sciences , cilt 12, (2022) Kasabov, N., Scott, N.M., Tu, E., Marks, S., Sengupta, N., Capecci, E., Othman, M., Doborjeh, M.G., Murli, N.: Hartono ve others, «Evolving spatio-temporal data machines based on the NeuCube neuromorphic framework: Design methodology and selected applications». Neural Networks cilt. 78 , 1–14 (2016) Olshausen ve, B.A., Field, D.J.: «Sparse coding of sensory inputs,» Current opinion in neurobiology , cilt 14, pp. 481–487, (2004) Hough, M., De Garis, H., Korkin, M., Gers ve, F., Nawa, N.E.: «SPIKER: Analog waveform to digital spiketrain conversion in ATR’s artificial brain (cam-brain) project,» %1 içinde International conference on robotics and artificial life , (1999) Schrauwen ve, B., Van Campenhout, J.: «BSA, a fast and accurate spike train encoding scheme,» %1 içinde Proceedings of the International Joint Conference on Neural Networks ,., 2003. (2003) Zeldenrust, F.: Wadman ve B. Englitz, «Neural coding with bursts—current state and future perspectives». Frontiers Comput. Neurosci. cilt. 12 , 48 (2018) Bonilla, L., Gautrais, J., Thorpe ve, S., Masquelier, T.: «Analyzing time-to-first-spike coding schemes: A theoretical approach». Frontiers Neurosci. cilt. 16 , 971937 (2022) Kayser, C., Montemurro, M.A., Logothetis, N.K. ve, Panzeri, S.: «Spike-phase coding boosts and stabilizes information carried by spatial and temporal spike patterns,» Neuron , cilt 61, pp. 597–608, (2009) Thorpe ve, S., Gautrais, J.: «Rank order coding,» %1 içinde. In: Computational Neuroscience: Trends in Research, 1998, pp. 113–118. Springer (1998) Averbeck, B.B., Latham, P.E. ve, Pouget, A.: «Neural correlations, population coding and computation,» Nature reviews neuroscience , cilt 7, pp. 358–366, (2006) Röth, K., Shao, S. ve, Gjorgjieva, J.: «Efficient population coding depends on stimulus convergence and source of noise,» PLoS computational biology , cilt 17, p. e1008897, (2021) Panzeri, S., Harvey, C.D., Piasini, E.: P. E. Latham ve T. Fellin, «Cracking the neural code for sensory perception by combining statistics, intervention, and behavior,» Neuron , cilt 93, pp. 491–507, (2017) Pachitariu, M., Lyamzin, D.R., Sahani ve, M., Lesica, N.A.: «State-dependent population coding in primary auditory cortex». Journal Neurosci. cilt. 35 , 2058–2073 (2015) Stringer, C., Pachitariu, M., Steinmetz, N.: Carandini ve K. D. Harris, «High-dimensional geometry of population responses in visual cortex». Nature cilt. 571 , 361–365 (2019) Panzeri, S., Moroni, M., Safaai, H. ve, Harvey, C.D.: «The structures and functions of correlations in neural population codes,» Nature Reviews Neuroscience , cilt 23, pp. 551–567, (2022) Wang, X., Lin ve, X., Dang, X.: «Supervised learning in spiking neural networks: A review of algorithms and evaluations,» Neural Networks , cilt 125, pp. 258–280, (2020) Bohte, S.M.: J. N. Kok ve H. La Poutre, «Error-backpropagation in temporally encoded networks of spiking neurons,» Neurocomputing , cilt 48, pp. 17–37, (2002) Tavanaei, A., Ghodrati, M., Kheradpisheh, S.R.: T. Masquelier ve A. Maida, «Deep learning in spiking neural networks,» Neural networks , cilt 111, pp. 47–63, (2019) Ponulak ve, F., Kasiński, A.: «Supervised learning in spiking neural networks with ReSuMe: sequence learning, classification, and spike shifting,» Neural computation , cilt 22, pp. 467–510, (2010) Gütig, R. ve, Sompolinsky, H.: «The tempotron: a neuron that learns spike timing–based decisions,» Nature neuroscience , cilt 9, pp. 420–428, (2006) Legenstein, R., Naeger, C. ve, Maass, W.: «What can a neuron learn with spike-timing-dependent plasticity?» Neural computation , cilt 17, pp. 2337–2382, (2005) Zenke ve, F., Ganguli, S.: «Superspike: Supervised learning in multilayer spiking neural networks,» Neural computation , cilt 30, pp. 1514–1541, (2018) Kheradpisheh ve, S.R., Masquelier, T.: «Temporal backpropagation for spiking neural networks with one spike per neuron». International J. neural Syst. cilt. 30 , 2050027 (2020) Liu, S., Leung ve, V.C.H., Dragotti, P.L.: «First-spike coding promotes accurate and efficient spiking neural networks for discrete events with rich temporal structures». Frontiers Neurosci. cilt. 17 , 1266003 (2023) Tang, G., Kumar, N., Yoo ve, R., Michmizos, K.: «Deep reinforcement learning with population-coded spiking neural network for continuous control,» %1 içinde Conference on Robot Learning , (2021) Moldwin ve, T., Segev, I.: «Perceptron Learning and Classification in a Modeled Cortical Pyramidal Cell». Frontiers Comput. Neurosci. cilt. 14 , 33 (April 2020) Amit, D.J., Wong, K.M. ve, Campbell, C.: «Perceptron learning with sign-constrained weights,» Journal of Physics A: Mathematical and General , cilt 22, p. 1989. (2039) Kayikcioglu Bozkir, I., Ozcan, Z., Kose, C., Kayikcioglu ve, T., Cetin, A.E.: «Improving a cortical pyramidal neuron model’s classification performance on a real-world ecg dataset by extending inputs». Journal Comput. Neurosci. cilt. 51 , 329–341 (2023) Fisher, R.A.: «The Use Of Multiple Measurements In Taxonomic Problems,» Annals of Eugenics , cilt 7, pp. 179–188, September (1936) Deng, L.: «The MNIST Database of Handwritten Digit Images for Machine Learning Research [Best of the Web],» IEEE Signal Processing Magazine , cilt 29, pp. 141–142, November (2012) Kramer, O.: «Scikit-Learn,» %1 içinde Machine Learning for Evolution Strategies , pp. 45–53. Springer International Publishing (2016) Hodgkin ve, A.L., Huxley, A.F.: «A quantitative description of membrane current and its application to conduction and excitation in nerve». Bulletin Math. Biology cilt. 52 , 25–71 (1952) Lapicque, L.M.: «Recherches quantitatives sur l’excitation electrique des nerfs». J Physiol. Paris cilt. 9 , 620–635 (1907) Izhikevich, E.M.: Dynamical Systems in Neuroscience: The Geometry of Excitability and Bursting. The MIT Press (2006) H. W. Kessels ve R. Malinow, «Synaptic AMPA Receptor Plasticity and Behavior,» Neuron , cilt 61, pp. 340–350, (2009) Mody, I., De Koninck, Y.: T. S. Otis ve I. Soltesz, «Bridging the cleft at GABA synapses in the brain,» Trends in Neurosciences , cilt 17, pp. 517–525, January (1994) Cover ve, T., Hart, P.: «Nearest neighbor pattern classification,» IEEE Transactions on Information Theory , cilt 13, pp. 21–27, January (1967) Cortes ve, C., Vapnik, V.: «Support-vector networks,» Machine Learning , cilt 20, pp. 273–297, September (1995) Cheeseman, P.C., Self, M., Kelly, J., Taylor, W.: D. Freeman ve J. C. Stutz, «Bayesian Classification.,» %1 içinde AAAI , (1988) Xanthopoulos, P., Pardalos, P.M., Trafalis, T.B., Xanthopoulos, P.: P. M. Pardalos ve T. B. Trafalis, «Linear discriminant analysis,» Robust data mining , p. 27–33, (2013) Rosenblatt, F.: «The perceptron: A probabilistic model for information storage and organization in the brain». Psychological Rev. cilt. 65 , 386–408 (1958) Hines ve, M., Carnevale, T.: «NEURON Simulation Environment,» %1 içinde Encyclopedia of Computational Neuroscience , cilt 9, Springer New York, pp. 2012–2017. (2015) Additional Declarations No competing interests reported. Supplementary Files Supplementary.rar Cite Share Download PDF Status: Under Revision Version 1 posted Editorial decision: Revision requested 29 May, 2025 Reviewers invited by journal 29 May, 2025 Editor assigned by journal 22 May, 2025 Submission checks completed at journal 22 May, 2025 First submitted to journal 20 May, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6709572","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":463732571,"identity":"d8cc7852-5258-44fc-98d8-4fc58739c60d","order_by":0,"name":"Zübeyir Özcan","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA9klEQVRIiWNgGAWjYFACHiA2gLI/HABTBrjUYmphnHHAAKolgZAWKGDmIUaL/Izcg495Cu7JybcffvbY5syfxAb25m0SjD/u4dTCOCMv2ZjHoNiYsSfN3DjnhkFiA8+xMgmGhGKcWpglcsykeQwSEpslGMykcz4AtQBFgFpwu4xNIsf8N0hLmwT7N2kLkBb5N/i18ADNZAZp6ZHgMZNmADkMyMCrRYLnXbLkHIMEYwmenDLJnjPGxm08acUWCWm4tci35x788OZPAjDEjm+T+HFMTraf/fDGGx9scGsBASYeZB4biMCvARjQPwgoGAWjYBSMghEOAEstSSYLoqVnAAAAAElFTkSuQmCC","orcid":"","institution":"Karadeniz Technical University","correspondingAuthor":true,"prefix":"","firstName":"Zübeyir","middleName":"","lastName":"Özcan","suffix":""},{"id":463732572,"identity":"d41429c4-e233-4bbb-9fd6-eb41075164fb","order_by":1,"name":"İlknur Kayıkçıoğlu Bozkır","email":"","orcid":"","institution":"Gümüşhane University","correspondingAuthor":false,"prefix":"","firstName":"İlknur","middleName":"Kayıkçıoğlu","lastName":"Bozkır","suffix":""},{"id":463732573,"identity":"f3b2e43e-12ff-4cc4-a61d-d1df5467d24a","order_by":2,"name":"Temel Kayıkçıoğlu","email":"","orcid":"","institution":"Karadeniz Technical University","correspondingAuthor":false,"prefix":"","firstName":"Temel","middleName":"","lastName":"Kayıkçıoğlu","suffix":""}],"badges":[],"createdAt":"2025-05-20 16:08:20","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6709572/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6709572/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":83764187,"identity":"ca8b1ad5-c2fa-4cf4-a4ef-1280dd9cfec1","added_by":"auto","created_at":"2025-06-02 10:33:28","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":188526,"visible":true,"origin":"","legend":"\u003cp\u003eVisual representation of the methodological pipeline, including neural \u0026nbsp;\u0026nbsp;network model configuration, spike pattern generation, and classification \u0026nbsp;\u0026nbsp;processes.\u003c/p\u003e","description":"","filename":"image1.png","url":"https://assets-eu.researchsquare.com/files/rs-6709572/v1/e77660ee2afe81ef51e9064e.png"},{"id":83763991,"identity":"d9fb7261-1bc3-41be-8398-e23f8823dfc1","added_by":"auto","created_at":"2025-06-02 10:25:28","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":109557,"visible":true,"origin":"","legend":"\u003cp\u003eVisualization of the spike pattern generation process using different \u0026nbsp;\u0026nbsp;encoding schemes. (A) Original continuous-valued feature vector and its quantized version. The rest of the rows represent spike patterns generated \u0026nbsp;\u0026nbsp;using Gray encoding (B), standard binary encoding (C) and intensity encoding (D).\u003c/p\u003e","description":"","filename":"image2.png","url":"https://assets-eu.researchsquare.com/files/rs-6709572/v1/38541328f45f88516a94ea4e.png"},{"id":83763992,"identity":"9d622bf1-f672-4c1c-b6e3-73e1b9368fea","added_by":"auto","created_at":"2025-06-02 10:25:28","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":131509,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 4. Decision boundaries obtained using the SLP \u0026nbsp;classifier on the synthetic dataset. Each subplot shows the classification regions for a specific combination of encoding scheme and quantization method. The “Analog” column represents the baseline result using unencoded feature vectors. Cluster labels (C1–C8) are indicated by color-coded regions.\u003c/p\u003e","description":"","filename":"image3.png","url":"https://assets-eu.researchsquare.com/files/rs-6709572/v1/d2bf9d50bbbade09b7d9a5df.png"},{"id":83763994,"identity":"b2c42993-9ff1-4136-bf26-fe398fbc2bac","added_by":"auto","created_at":"2025-06-02 10:25:28","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":105478,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 5. Decision boundaries obtained from the IZH model on the synthetic dataset. Each subplot represents the classification results for a specific combination of encoding method and quantization strategy. Cluster labels (C1–C8) are color-coded and plotted to visualize class separability in \u0026nbsp;\u0026nbsp;the encoded spike space.\u003c/p\u003e","description":"","filename":"image4.png","url":"https://assets-eu.researchsquare.com/files/rs-6709572/v1/da1c0664e90ffb004e59cf46.png"},{"id":83763995,"identity":"5c3d6696-3a58-4c54-9ed7-24b07270fad4","added_by":"auto","created_at":"2025-06-02 10:25:29","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":160805,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 6. Comparison of decision boundaries across different \u0026nbsp;\u0026nbsp;bit lengths and neuron models on the synthetic dataset. Each subplot \u0026nbsp;\u0026nbsp;illustrates the classification result for a specific bit length using \u0026nbsp;\u0026nbsp;intensity encoding and uniform quantization. The \"Analog\" column \u0026nbsp;\u0026nbsp;(far right) shows the baseline classification using raw feature vectors. \u0026nbsp;\u0026nbsp;Color-coded regions represent the predicted class areas for each cluster \u0026nbsp;\u0026nbsp;(C1–C8).\u003c/p\u003e","description":"","filename":"image5.png","url":"https://assets-eu.researchsquare.com/files/rs-6709572/v1/ee3cc00d778733f9f0f879d2.png"},{"id":83764731,"identity":"ae6dbdb2-c638-455c-a0f1-2ef271646ea0","added_by":"auto","created_at":"2025-06-02 10:41:30","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1460096,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6709572/v1/2699998b-5112-4957-95a9-868b0fc963d9.pdf"},{"id":83763998,"identity":"77d555ba-823c-440c-8626-afbb7de96461","added_by":"auto","created_at":"2025-06-02 10:25:29","extension":"rar","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":15596563,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementary.rar","url":"https://assets-eu.researchsquare.com/files/rs-6709572/v1/215844837c47d15b6d5be2a9.rar"}],"financialInterests":"No competing interests reported.","formattedTitle":"SPIGEN: A Configurable Population-Coded Spike Pattern Generator for Single-Spike Learning in Spiking Neural Networks","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eThe information processing capabilities of the brain are primarily realized through a vast network of interconnected neurons that communicate using spikes. Each neuron receives input in the form of structured spatiotemporal patterns transmitted via synaptic connections, processes this information according to its internal dynamics, and relays output to downstream neurons [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Due to its computational speed, sparsity, and energy efficiency, this mechanism has drawn significant attention from researchers in neuroscience, artificial intelligence, and computer science [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Attempts to emulate this form of computation have led to the development of biologically inspired models known as Spiking Neural Networks (SNNs). Over the past two decades, SNNs have found increasing applications in various domains, including neuromorphic hardware, robotics, and real-time sensory data processing [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eAlthough spiking neurons operate with spikes, real-world inputs are continuous. Biological systems resolve this via sensory organs, while artificial systems requires spike encoding to fill in this gap. As a result, various biologically inspired methods have been developed to convert continuous data into spike-based representations [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Broadly, these methods fall into two main categories: rate-based and temporal encoding strategies. Rate coding conveys information through the number of spikes emitted over a fixed time window, and it includes variants such as count rate coding and density rate coding. In contrast, temporal coding relies on the precise timing of spikes to represent stimulus characteristics. Several subclasses of temporal coding have been proposed in the literature, including temporal contrast [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e], correlation [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e], deconvolution [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e], latency/ISI [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e] and global reference [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e] based approaches. These diverse encoding paradigms have been applied across a range of SNN tasks, including sensory signal representation, temporal pattern recognition, and real-time event-driven neuromorphic computation [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eIn both rate and temporal coding strategies, continuous-valued stimuli are converted into spike trains by individual neurons, each encoding specific features of the input. In contrast, a third approach known as population coding represents stimuli through the collective activity of groups of neurons [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. In this approach, real-world stimuli are transformed into meaningful spatiotemporal spike patterns by multiple neurons whose dynamics are selectively tuned to different aspects of the input. Unlike previously mentioned encoding schemes, population coding distributes information across multiple synaptic channels which allows a richer and more robust representation [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. Recent studies have provided substantial evidence that information is represented through the coordinated activity of neuronal populations across various brain regions. Panzeri et al. [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e] showed that population coding happens at different levels in the brain. In the auditory cortex, Pachitariu et al. [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e] showed that how neuron populations respond to sounds depends on the brain\u0026rsquo;s state. Stringer et al. [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e] studied how groups of neurons respond to visual input in the mouse brain. Panzeri et al. [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e] studied how noise affects groups of neurons. They found that the amount of noise has an important impact on how well the brain can use population coding to carry information.\u003c/p\u003e \u003cp\u003eIn addition to spike encoding, the learning phase determines how synaptic weights are adjusted and governs the functional behavior of SNNs in performing various tasks. A wide range of learning approaches has been proposed for training SNNs. These methods can be divided into different subgroups based on network type, how learning is done, how information is encoded, how the structure changes, and how knowledge is represented [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. One of the earliest supervised learning methods, SpikeProp, adapts the classical backpropagation algorithm to spiking neurons with single spike outputs [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e], leading to multiple variants like QuickProp and Multi-SpikeProp [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. ReSuMe, combines STDP and anti-STDP to adjust synaptic weights based on mismatch between desired and actual spike trains [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. The Tempotron algorithm updates weights based on postsynaptic potential thresholds by capturing essential aspects of spike timing [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]. Reward-modulated STDP introduces global modulatory signals, mimicking dopamine-like mechanisms to guide synaptic plasticity based on delayed rewards [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. SuperSpike allows gradient-based training in multilayer SNNs by approximating the non-differentiable spike function with smooth surrogate functions [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eIn single-spike learning paradigms, where each neuron receives at most one spike per synapse, spike pattern generators are essential in encoding continuous inputs into sparse but informative spatiotemporal spike patterns. Several recent studies have combined population coding with single-spike learning to address this challenge. Kheradpisheh et al. [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e] proposed a rank-order coding strategy in which spike order encodes the input. Similarly, Liu et al. [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e] employed a first-spike decision model, and Tang et al. [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e] applied structured population codes in reinforcement learning. Moldwin and Segev [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e] demonstrated that binary input patterns can be successfully learned and classified by a biologically detailed pyramidal neuron. They adopted a biologically motivated constraint suggested by Amit et al. [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e], in which all synaptic weights are constrained to be positive-valued. As the input patterns were inherently binary, the system did not require a separate spike pattern generator. Their results showed that, despite the complexity and biophysical constraints of real neurons, such binary-coded inputs could be effectively separated through synaptic weight adaptation. A subsequent study by Kayık\u0026ccedil;ıoğlu Bozkır et al. [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e] adopted the same perceptron-based learning mechanism and investigated the classification performance of a pyramidal neuron model on real-world ECG data. They used Gray coding to convert analog ECG signals into binary spike patterns and assumed that all synaptic connections were excitatory. Their proposed mirroring approach led to a significant improvement in classification accuracy. Nevertheless, they did not investigate the potential impact of using alternative input encoding strategies in place of Gray coding on classification outcomes.\u003c/p\u003e \u003cp\u003eThis study proposes SPIGEN, a spike pattern generator, which addresses the challenge of encoding continuous-valued features for use in spiking neural networks. Unlike conventional rate or temporal encoding methods, the proposed approach transforms analog features into fixed-length binary spike patterns using a configurable pipeline consisting of quantization, coding scheme, and bit-length parameters. The main contributions of this study can be summarized as follows:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eIt introduces a flexible spike encoding architecture based on population coding, which supports multiple quantization and encoding schemes.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eIt provides a comprehensive evaluation of three quantization methods combined with three encoding strategies and bit length.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eIt analyzes how spike-based representations interact with both conventional classifiers and biologically inspired neuron models\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eIt incorporates both excitatory and inhibitory synaptic connections in the learning process, enhancing the biological plausibility.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e"},{"header":"2. Materials and Method","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1. Datasets\u003c/h2\u003e \u003cp\u003eThree different datasets were used in this study. Each dataset had a specific purpose. The IRIS and Digits datasets were used as standard benchmarks [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]. These datasets helped evaluate classification performance under different encoding strategies. Additionally, a synthetic Gaussian dataset was used for visualization. It allowed the examination of spatial structure and class separability in two dimensions.\u003c/p\u003e \u003cp\u003eAs the first dataset, the IRIS dataset includes 150 samples where each has four continuous features. The samples are grouped into three classes based on iris species [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e]. Secondly, The Digits dataset was taken from the scikit-learn library. It includes 1797 grayscale images of handwritten digits from 0 to 9. Each image is converted into a 64-dimensional feature vector [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]. Lastly, a synthetic dataset was created using the make_blobs function from scikit-learn [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]. It contains 1600 samples in two dimensions. The samples are grouped into eight Gaussian clusters. Each cluster has some overlap with others, controlled by a standard deviation of 0.6.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2. Methods\u003c/h2\u003e \u003cp\u003eThis study employes a computational method for classifying continuous-valued feature vectors using a spiking neural network. The method includes spike-based encoding, neural modeling, and supervised learning in a single system. Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e shows an overview of the full process.\u003c/p\u003e \u003cp\u003eIn the feature-to-spike pattern transformation phase, continuous-valued feature vectors are converted into spike patterns that spiking neural networks can process. This is done using SPIGEN, a spike pattern generator developed specifically for this study. The generator is trained using the statistical properties of the training data. After training, the same generator is used for both training and test data. In the network design phase, a spiking neural network is built to respond to spike patterns with class-specific outputs. The network has one layer of output neurons, and each neuron represents a class. Each bit in the spike pattern connects to all output neurons through two synapses: one excitatory and one inhibitory. The spike pattern acts directly as input to the output neurons through these synaptic connections. In the training and testing phase, the network is trained and tested using spike patterns created by SPIGEN. A supervised learning rule updates the synaptic weights during training. The synaptic weights are fixed once the training process is completed. In the testing stage, new spike patterns are given to the network using the same simulation settings.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cdiv id=\"Sec5\" class=\"Section3\"\u003e \u003ch2\u003e2.2.1. SPIGEN\u003c/h2\u003e \u003cp\u003eThis study proposes a method to generate spike patterns using population coding. The method converts continuous-valued feature vectors into spike patterns that can be interpreted by SNNs. It creates bit-wise spike patterns with fixed length. Each bit shows whether a spike is present, without depending on time. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e shows an example from the spike pattern generation process.\u003c/p\u003e \u003cp\u003eThe transformation begins with quantization. In this step, continuous-valued feature vectors are turned into a limited number of levels. The number of levels is set by the chosen bit length. To handle different data distributions, three quantization methods are used. Uniform quantization divides the feature range into equal intervals. Quantile-based quantization creates intervals with equal numbers of samples. K-means-based quantization groups values into clusters and assigns each to a level.\u003c/p\u003e \u003cp\u003eAfter quantization, each level is turned into a spike code. Three encoding methods are used. In standard binary encoding, the level is written in standard binary form. In Gray encoding, the level is written so that only one bit changes between two nearby levels. In intensity encoding, a number of bits at the beginning are set to one. The rest of the bits are zero. This way, the number of spikes corresponds to the level of the quantized feature. The spike codes from all features are flattened to create a single spike pattern for each sample. These patterns are simple in structure but still carry the important information from the original features.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section3\"\u003e \u003ch2\u003e2.2.2. Classfiers and Spiking Neural Network\u003c/h2\u003e \u003cp\u003eIn the study, two types of classification approaches were used to evaluate the performance of SPIGEN. In the first group, conventional machine learning algorithms were applied to both feature vectors and their corresponding spike patterns. These classifiers included k-Nearest Neigbors (k-NN) [\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e], Support Vector Machine (SVM) [\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e], Naive Bayes (NB) [\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e], Linear Discriminant Analysis (LDA) [\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e], and Single Layer Perceptron (SLP) [\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e], all implemented using the scikit-learn library with default settings.\u003c/p\u003e \u003cp\u003eIn the second group, SNNs were used to test how well the spike patterns support classification. The networks had only one output layer for each. Each neuron in the output layer represented a class. Each input sample is given to the network as a spike pattern. Each bit in the pattern acts as a presynaptic spike. These spikes are sent through fixed synaptic connections to all output neurons at the same time. To see how neuronal dynamics affect spike-based pattern processing, three neuron models were tested. Each model was used in a separate version of the same network. The input encoding, synaptic setup, and training method stayed the same for all models.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section3\"\u003e \u003ch2\u003e2.2.3. Neuron Models and Synaptic Connections\u003c/h2\u003e \u003cp\u003eWhen building SNNs, three biologically inspired spiking neuron models which are Hodgkin-Huxley (HH) [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e], Leaky Integrate-and-Fire (LIF) [\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e], and Izhikevich (IZH) [\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e] were utilized. Each model was tested separately within the same network structure. The input encoding, synaptic setup, and learning procedure were kept the same across all cases. This allowed a fair comparison of how each neuron\u0026rsquo;s internal behavior affects classification.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eModel parameters used in simulations. The table lists the biophysical parameters for the LIF, IZH, and HH neuron models, as well as the synaptic properties of AMPA and GABA synapses.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eParameter\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eValue\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eParameter\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eValue\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"5\" rowspan=\"6\"\u003e \u003cp\u003eLIF\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eV\u003csub\u003erest\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-60 mV\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\" morerows=\"8\" rowspan=\"9\"\u003e \u003cp\u003eIZH\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eV\u003csub\u003erest\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e-60 mV\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eV\u003csub\u003ethreshold\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-45 mV\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eV\u003csub\u003ethreshold\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e-40 mV\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eV\u003csub\u003ereset\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-60 mV\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eV\u003csub\u003epeak\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e35 mV\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eV\u003csub\u003epeak\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e35 mV\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eC\u003csub\u003em\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1 mF\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eC\u003csub\u003em\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.45 mF\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ek\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.7 nS/mV\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eg\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.001 \u0026micro;S\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ea\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.03 1/ms\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"6\" rowspan=\"7\"\u003e \u003cp\u003eHH\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eE\u003csub\u003eNa\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e55.17 mV\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eb\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e-2 nS\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eE\u003csub\u003eK\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-72.14 mV\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ec\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e-50 mV\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eE\u003csub\u003eleak\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-49.42 mV\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ed\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e100 pA\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eg\u003csub\u003eNa\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.2 \u0026micro;S\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eAMPA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eE\u003csub\u003eAMPA\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0 mV\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eg\u003csub\u003eK\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.36 \u0026micro;S\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eτ\u003csub\u003eAMPA\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e5 ms\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eg\u003csub\u003eleak\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.003 \u0026micro;S\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eGABA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eE\u003csub\u003eGABA\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e-70 mV\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eC\u003csub\u003em\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.01 mF\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eτ\u003csub\u003eGABA\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e6 ms\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eIn this study, AMPA synapses were used to model excitatory connections [\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e], while GABA\u003csub\u003eA\u003c/sub\u003e synapses were used to represent inhibitory ones [\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e]. These synaptic types were chosen for their biological relevance and fast response characteristics. The specific parameter settings used for each model are summarized in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Synaptic weights were initially set to zero, assuming no functional connection at the beginning of training. These values were then updated during an adapted perceptron rule.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section3\"\u003e \u003ch2\u003e2.2.4. Learning Rule\u003c/h2\u003e \u003cp\u003eIn this study, an adapted version of the classical perceptron learning rule is employed to update synaptic weights [\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e]. In classical rule, the weight update is performed as follows:\u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e\n$$\\:{{w}_{i}}^{New}\\leftarrow\\:{{w}_{i}}^{Old}+\\varDelta\\:{w}_{i}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equ2\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ2\" name=\"EquationSource\"\u003e\n$$\\:\\varDelta\\:{w}_{i}=\\eta\\:\\times\\:\\left(d-out\\right)\\times\\:{x}_{i}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e2\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eHere, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:out\\)\u003c/span\u003e\u003c/span\u003e refers to the binary output of the neuron, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:d\\)\u003c/span\u003e\u003c/span\u003e is the desired output, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{x}_{i}\\)\u003c/span\u003e\u003c/span\u003e is the input at position \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:i\\)\u003c/span\u003e\u003c/span\u003e, and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\varDelta\\:{w}_{i}\\)\u003c/span\u003e\u003c/span\u003e denotes the change in the synaptic weight. According to this rule, the weights are updated proportionally to the classification error and the input activation.\u003c/p\u003e \u003cp\u003eWhile this rule is applicable to artificial neurons without constraints on weight values, it requires modification for use with spiking neurons, especially in biologically inspired settings. Moldwin and Segev [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e] highlighted that synaptic weights in biological neurons must remain non-negative, since negative values would imply non-physical conductance values. To enforce this constraint, they proposed a modified update rule. This rule ensures that if a weight becomes negative during training, it is instead clipped to zero. However, this formulation only applies to excitatory synapses and does not address inhibitory ones. Since the present study includes both excitatory and inhibitory synapses, Eq.\u0026nbsp;\u003cspan refid=\"Equ2\" class=\"InternalRef\"\u003e2\u003c/span\u003e is extended to account for synaptic polarity and weight constraint. The generalized update rule is defined as follows:\u003cdiv id=\"Equ3\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ3\" name=\"EquationSource\"\u003e\n$$\\:{{w}_{i}}^{New}\\leftarrow\\:\\left\\{\\begin{array}{cc}max\\left(0,\\left({{w}_{i}}^{Old}+\\varDelta\\:{w}_{i}\\right)\\right),\u0026amp;\\:For\\:excitatory\\:synapse\\\\\\:max\\left(0,\\left({{w}_{i}}^{Old}-\\varDelta\\:{w}_{i}\\right)\\right),\u0026amp;\\:For\\:inhibitory\\:synapse\\end{array}\\right.$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e3\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eEquation \u003cspan refid=\"Equ3\" class=\"InternalRef\"\u003e3\u003c/span\u003e allows the learning rule to update weights while preserving the non-negativity and polarity constraints of both synapse types. Additionally, for constant input terms such as \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{I}_{base}\\)\u003c/span\u003e\u003c/span\u003e​, weight updates are applied according to Eq.\u0026nbsp;\u003cspan refid=\"Equ2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, as in standard perceptron learning.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section3\"\u003e \u003ch2\u003e2.2.5. Training and Testing\u003c/h2\u003e \u003cp\u003eThe training process begins by selecting a training sample and converting it into a spike pattern using the previously constructed spike pattern generator. Each input in the pattern is associated with two predefined synaptic connections which are excitatory and inhibitory. The spiking neural network is then simulated for 50 ms. At the 10th ms of the simulation, input spikes are applied to synaptic connections corresponding to the inputs with a value of 1 in the spike pattern. During the simulation, the membrane potentials of all output neurons are recorded. An output neuron is labeled as active if it produces at least one spike, and inactive otherwise. If an output neuron fails to spike when it should, or spikes when it should not, its synaptic weights are updated using the learning rule. After processing a training sample, the procedure continues with the next one until all samples are used. One complete pass through all training samples is considered an epoch. If the network still produces incorrect outputs, training continues for additional epochs until the desired accuracy is achieved or a maximum epoch limit is reached. At the end of training, the weights are fixed and the system proceeds to the testing phase.\u003c/p\u003e \u003cp\u003eTesting also consists of two stage which are simulation and decision. For each test sample, a spike pattern is generated using the same spike pattern generator used in training. The pattern is mapped onto the network using the same configuration and synaptic connections. Class labels are decided based on the outputs of the neurons that detect specific classes. For each test spike pattern, three different cases may happen. In the first case, only one output neuron produces spikes where the other neurons stay silent. In this case, the class of the active neuron is chosen as the label. In the second case, more than one output neuron produces spikes. Then, the label is taken from the neuron that spikes first. In the third case, no output neuron spikes. In this situation, membrane potentials are checked. The label is given based on the neuron that has the highest membrane potential.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e2.3. Experimental Design and Simulation Environment\u003c/h2\u003e \u003cp\u003eThe experimental procedure was carried out in two main stages. In the first stage, classical machine learning algorithms were applied directly to the raw analog feature vectors to evaluate their baseline classification performance. In the case of synthetic dataset, decision boundaries were visualized in two-dimensional space. For IRIS and Digits datasets, performance was evaluated using the F1-score. No spiking neuron models were used in this first stage.\u003c/p\u003e \u003cp\u003eIn the second stage, all datasets were passed through the spike pattern generation process. The resulting spike patterns were evaluated using the same classification methods as in the first stage. In addition, three spiking neuron models were included as SNN classifiers. For all spiking models, the number of training epochs was set to 500. For the spike pattern generator, bit length was selected as 5, quantization was selected as uniform, and encoding was selected as intensity coding by default. To assess the effect of encoding parameters on classification performance, the influence of bit length, quantization strategy, and encoding scheme was evaluated independently. In the case of the Digits dataset, which is already quantized into 16 levels, experiments related to quantization and bit length were excluded.\u003c/p\u003e \u003cp\u003eFor the IRIS and Digits datasets, the original data were shuffled and then split into two parts. 70% of the data were used for training, and 30% were used for testing. This process was repeated 5 times to get more reliable results. Each time, the data were shuffled and split again. The results from each shuffle were calculated separately. In the end, the average of all results was taken for evaluation.\u003c/p\u003e \u003cp\u003eAll simulations involving spiking neuron models were conducted using the NEURON simulation environment, which provides biophysically realistic modeling capabilities for neuronal dynamics [\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e]. The simulator was integrated with Python for data preprocessing, spike pattern generation, parameter configuration, and result analysis. The experiments involving conventional machine learning algorithms were conducted using Python\u0026rsquo;s scikit-learn library [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]. The model specific parameters were chosen as the default configurations defined by the library.\u003c/p\u003e \u003c/div\u003e"},{"header":"3. Results","content":"\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e3.1. IRIS Dataset\u003c/h2\u003e \u003cp\u003eThe IRIS dataset was used to test how spike pattern generator configurations affect classification performance. Different types of classifiers were included in two separate experiments. In the first experiment, all encoding schemes and quantization methods were combined and tested while the bit length was kept as 5. In the second experiment, bit length was iterated from 2 to 7, while uniform quantization and intensity encoding were used. In both experiments, the F1-score was used to measure performance. Results were calculated for both classical classifiers and spiking neural network models and mapped on Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e and Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e shows the performance of classical machine learning models. Continuous features gave the highest f1-scores for all classifiers except for SLP. These scores were used as the performance benchmark. While SVM had the best score with 98.2%, LDA and k-NN classifiers followed with 99.1%, and 96.0%, respectively. Intensity encoding gave results close to the continuous baseline. This was clearer when it was used with uniform or quantile quantization. SLP showed high performance with intensity and Gray encoding, reaching up to 96.0%. NB has shown high sensitivity to encoding types with respect to baseline, especially when using binary encoding. With uniform quantization, its score even dropped to 80.5%.\u003c/p\u003e \u003cp\u003eResults from SNNs have shown that intensity encoding promises the most stable and high results. When used with uniform or quantile quantization, HH, IZH, and LIF had f1-scores over 94%. These results were close to the scores of classical classifiers. Gray encoding showed more variation. Its performance changed with different quantization methods. The scores were between 85.6% and 92.4%. Binary encoding gave the lowest scores for SNNs. In the HH model, its score dropped to 63.1% with k-means quantization.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eF1-score results (%) of IRIS dataset obtained using different combinations of encoding and quantization strategies across various classifiers. The \u0026ldquo;Continuous\u0026rdquo; row shows baseline results without encoding. Here the bit length was chosen as 5. Each value represents the average accuracy over five runs.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"10\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c10\" colnum=\"10\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCoding\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eQuantization\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eKNN\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSVM\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSLP\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNB\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eLDA\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eHH\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c9\"\u003e \u003cp\u003eIZH\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c10\"\u003e \u003cp\u003eLIF\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eContinuous\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eContinuous\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e96,0\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e98,2\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e92,0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e94,7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e\u003cb\u003e99,1\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eIntensity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eUniform\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e95,1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e94,7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e95,1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e94,7\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e87,9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e\u003cb\u003e95,1\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e\u003cb\u003e94,7\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e \u003cp\u003e\u003cb\u003e94,7\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eQuantile\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e95,6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e93,8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e96,0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e93,3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e89,9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e\u003cb\u003e95,1\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e93,8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e \u003cp\u003e\u003cb\u003e94,7\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ek-Means\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e95,6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e94,2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e95,1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e93,4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e82,9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e94,7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e93,8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e \u003cp\u003e94,7\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eGray\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eUniform\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e94,3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e93,3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e96,0\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e92,5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e94,3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e85,6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e92,4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e \u003cp\u003e93,8\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eQuantile\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e91,1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e94,2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e90,2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e90,8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e92,1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e74,0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e87,6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e \u003cp\u003e90,0\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ek-Means\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e92,9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e92,4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e90,7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e87,8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e93,4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e77,5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e89,2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e \u003cp\u003e90,2\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eBinary\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eUniform\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e84,1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e92,4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e74,9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e80,5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e95,1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e72,5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e83,1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e \u003cp\u003e83,6\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eQuantile\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e85,0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e93,8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e87,0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e85,3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e95,6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e66,4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e87,4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e \u003cp\u003e88,4\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ek-Means\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e81,5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e92,9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e80,2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e83,4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e95,6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e63,1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e84,4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e \u003cp\u003e85,3\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eF1-score results (%) of IRIS dataset obtained for different bit lengths used in the spike pattern generator. The \u0026ldquo;Continuous\u0026rdquo; row shows baseline results without encoding. Here, the quantization strategy and coding scheme were chosen as uniform and intensity, respectively.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"9\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBit Length\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eKNN\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSVM\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSLP\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eNB\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eLDA\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eHH\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eIZH\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c9\"\u003e \u003cp\u003eLIF\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eContinuous\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003e96,0\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e98,2\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e92,0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e94,7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e99,1\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e92,1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e95,1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e94,3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e87,0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e90,3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e88,2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e91,3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e91,3\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e94,2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e94,2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e94,7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e94,7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e94,2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e89,8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e93,4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e94,2\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e95,1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e95,6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e95,6\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e93,8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e94,6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e93,3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e93,8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e94,2\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e95,1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e94,7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e95,1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e94,7\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e87,9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e\u003cb\u003e95,1\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e\u003cb\u003e94,7\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e\u003cb\u003e94,7\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e95,1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e93,8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e95,1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e92,9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e74,9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e94,7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e93,8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e93,8\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e95,1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e93,8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e95,1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e93,4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e74,4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e94,7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e93,8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e94,2\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e shows that the encoding scheme and quantization method have a significant effect on classification performance. Intensity encoding gave better results than Gray and binary coding in most classifiers. Uniform and quantile quantization approaches gave higher f1-scores than k-means quantization. This was clearer when they were matched with intensity encoding. Binary encoding is simple, but it gave the lowest performance. This was more obvious when it was used with k-means quantization. Using intensity encoding with uniform or quantile quantization gives a strong combination for both traditional machine learning algorithms and spiking neural networks.\u003c/p\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e shows the f1-score results for different bit lengths in the spike pattern generator. As expected, increasing the bit length improved classification performance in most classifiers. But this improvement continued only up to a certain point. There was a clear increase from 2-bit to 4-bit lengths. This was more visible in SVM, NB, and the spiking neuron models. For example, the HH model increased from 88.2% at 2 bits to 93.3% at 4 bits. However, after 5 bits, performance stopped improving. In some cases, like LDA and NB, it even started to drop. According to table, choosing the bit length as 4 or 5 seems to give the best balance between computational load and classification performance.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e3.2. Digits Dataset\u003c/h2\u003e \u003cp\u003eThe Digits dataset was used to evaluate the impact of different encoding schemes on classification performance. Since the input feature vectors in this dataset are already quantized into 16 discrete levels, further quantization or bit length variation was not necessary. Therefore, the experiments focused on comparing the performance of different encoding schemes. Classification was performed using both traditional machine learning models and spiking neural networks models, and the F1-score was used as the evaluation metric.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eF1-score results (%) of Digits dataset for different encoding schemes applied to the same feature vectors. The \u0026ldquo;Continuous\u0026rdquo; row shows baseline results without encoding.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"9\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCoding\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eKNN\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSVM\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSLP\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eNB\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eLDA\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eHH\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eIZH\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c9\"\u003e \u003cp\u003eLIF\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eContinuous\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003e98,0\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e97,7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e95,3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e83,4\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e94,9\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIntensity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e97,6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e98,0\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e97,0\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e80,2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e87,7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e\u003cb\u003e95,6\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e\u003cb\u003e95,4\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e\u003cb\u003e95,6\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGray\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e88,7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e94,3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e91,5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e74,1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e92,1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e91,2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e91,2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e91,4\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBinary\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e92,8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e96,2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e94,4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e72,9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e93,9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e93,0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e93,2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e93,2\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e presents the f1-score results for the Digits dataset obtained using different encoding schemes. The continuous feature representation achieved the highest overall performance in most classical classifiers, particularly with SVM and k-NN. Among the encoded variants, intensity encoding consistently outperformed Gray and binary across both traditional and SNNs. With intensity encoding, SLP reached 97.0%, overcoming its baseline. The HH, IZH, and LIF models also achieved scores above 95%. Gray and binary encoding resulted in lower performance. These results confirm that intensity encoding is the most effective scheme for preserving the discriminative structure of the input space when working with pre-quantized features such as those in the Digits dataset.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e3.3. Synthetic Dataset\u003c/h2\u003e \u003cp\u003eThe synthetic dataset was used to visually analyze how different encoding schemes and parameter configurations affect class separability in both traditional and spiking neural networks. Its two-dimensional structure and clearly clustered class distribution make it suitable for illustrating decision boundaries. Three figures are presented in this section. The first shows the decision regions obtained using SLP classifier for all combinations of encoding schemes and quantization strategies. Although only SLP results are visualized, results for additional machine learning classifiers were also evaluated and are included in the supplementary material. The second figure illustrates the same encoding and quantization configurations applied to the IZH model. Similarly, results for other spiking models were also generated and are provided in the supplementary material. The final figure examines the effect of varying bit length on classification boundaries across neuron models using fixed intensity encoding and uniform quantization.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e4\u003c/span\u003e presents the decision boundaries obtained using the SLP for different combinations of encoding schemes and quantization methods. As expected, the continuous input yielded the cleanest and most compact class regions. Among the encoded inputs, intensity encoding produced the most coherent and stable decision boundaries, especially when paired with quantile or uniform quantization. Gray encoding resulted in moderately structured regions but exhibited irregular edges and fragmented zones. Binary encoding produced the noisiest and most scattered boundaries across all quantization methods. These results demonstrate that intensity encoding allows SLP to approximate continuous classification behavior more closely than other encoding strategies.\u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e5\u003c/span\u003e shows the decision boundaries obtained from the IZH model results. Among the three encoding types, intensity encoding once again demonstrates superior performance, producing smooth and well-structured decision regions. Gray encoding provides moderate separability but suffers from fragmented and fuzzy boundaries. Binary encoding shows the weakest boundary structure, with highly scattered and disorganized regions. Compared to the SLP results, the decision surfaces obtained from the IZH model are generally coarser and less defined. Nevertheless, the IZH model, when paired with intensity encoding and proper quantization, is capable of capturing meaningful spatial structure. These results confirm the compatibility of the proposed spike encoding method with spiking neural networks.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e6\u003c/span\u003e illustrates the effect of varying bit length on the decision boundaries generated by four different models. All configurations use intensity encoding and uniform quantization. Across all models, increasing the bit length generally improves the structure and clarity of the decision regions. With n\u0026thinsp;=\u0026thinsp;2, decision boundaries are noisy, where class regions often overlap. As the bit length increases to 4 or 5, boundaries become sharper and more aligned with cluster distributions. For most models, n\u0026thinsp;=\u0026thinsp;5 provides a good trade-off between resolution and generalization. Beyond 5 bits, the improvements tend to stop. These visualizations reinforce the findings from the IRIS dataset, showing that moderate bit lengths are sufficient to encode meaningful spike patterns across both conventional and SNNs.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"4. Discussion","content":"\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e4.1. Effect of Quantization\u003c/h2\u003e \u003cp\u003eIn the experiments, quantization strategy was found to have a significant impact on classification performance. Across both traditional classifiers and spiking neural networks, quantile and uniform quantization generally outperformed k-means clustering. This trend was evident in the IRIS dataset results, where intensity encoding paired with quantile or uniform quantization consistently achieved higher f1-scores across most models. Visual inspection of decision boundaries in the synthetic dataset further supports this finding. Class regions under quantile and uniform quantization appeared more structured and aligned with true class distributions. In contrast, k-means quantization led to irregular and fragmented boundaries. These results indicate that quantile and uniform quantization methods provide more stable and generalizable input representations.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e4.2. Effect of Coding Scheme\u003c/h2\u003e \u003cp\u003eThe choice of coding scheme had a pronounced effect on classification performance, both in terms of quantitative performance and decision boundary structure. Among the evaluated schemes, intensity encoding consistently delivered the highest f1-scores across datasets and classifier types. As shown in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e and Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, intensity encoding performed comparably to continuous representations. Visualizations from the synthetic dataset reinforced this trend. Intensity-encoded patterns produced well-separated class regions. Gray encoding demonstrated moderate performance, providing smoother transitions than binary but still suffering from instability under certain quantization methods. Binary encoding yielded the lowest performance overall, characterized by noisy and fragmented class boundaries. These results suggest that intensity encoding is more effective in preserving the structure of features. This makes it a robust scheme for spike patter generation.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003e4.3. Effect of Bit Length\u003c/h2\u003e \u003cp\u003eBit length, which determines the resolution of the spike pattern representation, was shown to influence classification performance non-linearly. Results from the IRIS dataset and the synthetic dataset revealed that increasing the bit length from 2 to 4 or 5 generally led to significant improvements in both f1-score and decision boundary clarity. This trend was observed across both traditional classifiers and SNNs. Notably, performance gains plateaued beyond 5 bits, and in some cases, further increases led to declines. These findings indicate that moderate bit lengths are sufficient for capturing class-discriminative information in spike pattern generation, while excessively high resolutions may introduce unnecessary complexity without tangible benefit.\u003c/p\u003e \u003c/div\u003e"},{"header":"5. Conclusion","content":"\u003cp\u003eThis study introduced SPIGEN, a spike pattern generation framework, designed to convert continuous-valued feature vectors into fixed-length, biologically interpretable spike patterns. The method enables compatibility with both traditional machine learning classifiers and SNNs by incorporating configurable quantization, encoding, and bit-length parameters. Experimental results across multiple datasets showed that intensity encoding combined with uniform or quantile quantization offers the most robust classification performance, closely approximating continuous feature baselines. Moderate bit lengths, particularly 4 and 5 bits, were found to provide an optimal balance between resolution and performance. Visual analyses on a synthetic dataset further confirmed that the proposed encoding strategy preserves spatial class structure. These findings highlight the effectiveness of SPIGEN and its suitability for neuromorphic computing, embedded learning systems, and low-power spike-based processing. Future work may explore adaptive encoding mechanisms, online learning extensions, and hardware implementations to further enhance practical applicability.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eFunding:\u003c/h2\u003e \u003cp\u003eThis research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eZ\u0026uuml;beyir \u0026Ouml;zcan: Conceptualization, Methodology, Software, Formal analysis, Investigation, Data curation, Visualization, Writing \u0026ndash; Original Draft.İlknur Kayık\u0026ccedil;ıoğlu Bozkır: Validation, Writing \u0026ndash; Review \u0026amp; Editing, Resources.Temel Kayık\u0026ccedil;ıoğlu: Supervision, Project administration, Writing \u0026ndash; Review \u0026amp; Editing.All authors have read and approved the final manuscript.\u003c/p\u003e\u003ch2\u003eData Availability:\u003c/h2\u003e \u003cp\u003eNo new data sets were acquired in this study. Code and trained models will be made available on reasonable request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eGerstner, W., Kistler, W.M., Naud ve, R.: L. Paninski, Neuronal Dynamics: From Single Neurons to Networks and Models of Cognition. Cambridge University Press (2014)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAuge, D., Hille, J., Mueller ve, E., Knoll, A.: \u0026laquo;A survey of encoding techniques for signal processing in spiking neural networks\u0026raquo;. Neural Process. Lett. cilt. \u003cb\u003e53\u003c/b\u003e, 4693\u0026ndash;4710 (2021)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eForno, E., Fra, V., Pignari, R., Macii ve, E., Urgese, G.: \u0026laquo;Spike encoding techniques for IoT time-varying signals benchmarked on a neuromorphic classification task\u0026raquo;. Frontiers Neurosci. cilt. \u003cb\u003e16\u003c/b\u003e, 999029 (2022)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYamazaki, K., Vo-Ho, V.-K.: D. Bulsara ve N. Le, \u0026laquo;Spiking Neural Networks and Their Applications: A Review,\u0026raquo; \u003cem\u003eBrain Sciences\u003c/em\u003e, cilt 12, (2022)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKasabov, N., Scott, N.M., Tu, E., Marks, S., Sengupta, N., Capecci, E., Othman, M., Doborjeh, M.G., Murli, N.: Hartono ve others, \u0026laquo;Evolving spatio-temporal data machines based on the NeuCube neuromorphic framework: Design methodology and selected applications\u0026raquo;. Neural Networks cilt. \u003cb\u003e78\u003c/b\u003e, 1\u0026ndash;14 (2016)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOlshausen ve, B.A., Field, D.J.: \u0026laquo;Sparse coding of sensory inputs,\u0026raquo; \u003cem\u003eCurrent opinion in neurobiology\u003c/em\u003e, cilt 14, pp. 481\u0026ndash;487, (2004)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHough, M., De Garis, H., Korkin, M., Gers ve, F., Nawa, N.E.: \u0026laquo;SPIKER: Analog waveform to digital spiketrain conversion in ATR\u0026rsquo;s artificial brain (cam-brain) project,\u0026raquo; %1 i\u0026ccedil;inde \u003cem\u003eInternational conference on robotics and artificial life\u003c/em\u003e, (1999)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchrauwen ve, B., Van Campenhout, J.: \u0026laquo;BSA, a fast and accurate spike train encoding scheme,\u0026raquo; %1 i\u0026ccedil;inde \u003cem\u003eProceedings of the International Joint Conference on Neural Networks\u003c/em\u003e,., 2003. (2003)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZeldenrust, F.: Wadman ve B. Englitz, \u0026laquo;Neural coding with bursts\u0026mdash;current state and future perspectives\u0026raquo;. Frontiers Comput. Neurosci. cilt. \u003cb\u003e12\u003c/b\u003e, 48 (2018)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBonilla, L., Gautrais, J., Thorpe ve, S., Masquelier, T.: \u0026laquo;Analyzing time-to-first-spike coding schemes: A theoretical approach\u0026raquo;. Frontiers Neurosci. cilt. \u003cb\u003e16\u003c/b\u003e, 971937 (2022)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKayser, C., Montemurro, M.A., Logothetis, N.K. ve, Panzeri, S.: \u0026laquo;Spike-phase coding boosts and stabilizes information carried by spatial and temporal spike patterns,\u0026raquo; \u003cem\u003eNeuron\u003c/em\u003e, cilt 61, pp. 597\u0026ndash;608, (2009)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eThorpe ve, S., Gautrais, J.: \u0026laquo;Rank order coding,\u0026raquo; %1 i\u0026ccedil;inde. In: Computational Neuroscience: Trends in Research, 1998, pp. 113\u0026ndash;118. Springer (1998)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAverbeck, B.B., Latham, P.E. ve, Pouget, A.: \u0026laquo;Neural correlations, population coding and computation,\u0026raquo; \u003cem\u003eNature reviews neuroscience\u003c/em\u003e, cilt 7, pp. 358\u0026ndash;366, (2006)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eR\u0026ouml;th, K., Shao, S. ve, Gjorgjieva, J.: \u0026laquo;Efficient population coding depends on stimulus convergence and source of noise,\u0026raquo; \u003cem\u003ePLoS computational biology\u003c/em\u003e, cilt 17, p. e1008897, (2021)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePanzeri, S., Harvey, C.D., Piasini, E.: P. E. Latham ve T. Fellin, \u0026laquo;Cracking the neural code for sensory perception by combining statistics, intervention, and behavior,\u0026raquo; \u003cem\u003eNeuron\u003c/em\u003e, cilt 93, pp. 491\u0026ndash;507, (2017)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePachitariu, M., Lyamzin, D.R., Sahani ve, M., Lesica, N.A.: \u0026laquo;State-dependent population coding in primary auditory cortex\u0026raquo;. Journal Neurosci. cilt. \u003cb\u003e35\u003c/b\u003e, 2058\u0026ndash;2073 (2015)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStringer, C., Pachitariu, M., Steinmetz, N.: Carandini ve K. D. Harris, \u0026laquo;High-dimensional geometry of population responses in visual cortex\u0026raquo;. Nature cilt. \u003cb\u003e571\u003c/b\u003e, 361\u0026ndash;365 (2019)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePanzeri, S., Moroni, M., Safaai, H. ve, Harvey, C.D.: \u0026laquo;The structures and functions of correlations in neural population codes,\u0026raquo; \u003cem\u003eNature Reviews Neuroscience\u003c/em\u003e, cilt 23, pp. 551\u0026ndash;567, (2022)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang, X., Lin ve, X., Dang, X.: \u0026laquo;Supervised learning in spiking neural networks: A review of algorithms and evaluations,\u0026raquo; \u003cem\u003eNeural Networks\u003c/em\u003e, cilt 125, pp. 258\u0026ndash;280, (2020)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBohte, S.M.: J. N. Kok ve H. La Poutre, \u0026laquo;Error-backpropagation in temporally encoded networks of spiking neurons,\u0026raquo; \u003cem\u003eNeurocomputing\u003c/em\u003e, cilt 48, pp. 17\u0026ndash;37, (2002)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTavanaei, A., Ghodrati, M., Kheradpisheh, S.R.: T. Masquelier ve A. Maida, \u0026laquo;Deep learning in spiking neural networks,\u0026raquo; \u003cem\u003eNeural networks\u003c/em\u003e, cilt 111, pp. 47\u0026ndash;63, (2019)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePonulak ve, F., Kasiński, A.: \u0026laquo;Supervised learning in spiking neural networks with ReSuMe: sequence learning, classification, and spike shifting,\u0026raquo; \u003cem\u003eNeural computation\u003c/em\u003e, cilt 22, pp. 467\u0026ndash;510, (2010)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eG\u0026uuml;tig, R. ve, Sompolinsky, H.: \u0026laquo;The tempotron: a neuron that learns spike timing\u0026ndash;based decisions,\u0026raquo; \u003cem\u003eNature neuroscience\u003c/em\u003e, cilt 9, pp. 420\u0026ndash;428, (2006)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLegenstein, R., Naeger, C. ve, Maass, W.: \u0026laquo;What can a neuron learn with spike-timing-dependent plasticity?\u0026raquo; \u003cem\u003eNeural computation\u003c/em\u003e, cilt 17, pp. 2337\u0026ndash;2382, (2005)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZenke ve, F., Ganguli, S.: \u0026laquo;Superspike: Supervised learning in multilayer spiking neural networks,\u0026raquo; \u003cem\u003eNeural computation\u003c/em\u003e, cilt 30, pp. 1514\u0026ndash;1541, (2018)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKheradpisheh ve, S.R., Masquelier, T.: \u0026laquo;Temporal backpropagation for spiking neural networks with one spike per neuron\u0026raquo;. International J. neural Syst. cilt. \u003cb\u003e30\u003c/b\u003e, 2050027 (2020)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu, S., Leung ve, V.C.H., Dragotti, P.L.: \u0026laquo;First-spike coding promotes accurate and efficient spiking neural networks for discrete events with rich temporal structures\u0026raquo;. Frontiers Neurosci. cilt. \u003cb\u003e17\u003c/b\u003e, 1266003 (2023)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTang, G., Kumar, N., Yoo ve, R., Michmizos, K.: \u0026laquo;Deep reinforcement learning with population-coded spiking neural network for continuous control,\u0026raquo; %1 i\u0026ccedil;inde \u003cem\u003eConference on Robot Learning\u003c/em\u003e, (2021)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMoldwin ve, T., Segev, I.: \u0026laquo;Perceptron Learning and Classification in a Modeled Cortical Pyramidal Cell\u0026raquo;. Frontiers Comput. Neurosci. cilt. \u003cb\u003e14\u003c/b\u003e, 33 (April 2020)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAmit, D.J., Wong, K.M. ve, Campbell, C.: \u0026laquo;Perceptron learning with sign-constrained weights,\u0026raquo; \u003cem\u003eJournal of Physics A: Mathematical and General\u003c/em\u003e, cilt 22, p. 1989. (2039)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKayikcioglu Bozkir, I., Ozcan, Z., Kose, C., Kayikcioglu ve, T., Cetin, A.E.: \u0026laquo;Improving a cortical pyramidal neuron model\u0026rsquo;s classification performance on a real-world ecg dataset by extending inputs\u0026raquo;. Journal Comput. Neurosci. cilt. \u003cb\u003e51\u003c/b\u003e, 329\u0026ndash;341 (2023)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFisher, R.A.: \u0026laquo;The Use Of Multiple Measurements In Taxonomic Problems,\u0026raquo; \u003cem\u003eAnnals of Eugenics\u003c/em\u003e, cilt 7, pp. 179\u0026ndash;188, September (1936)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDeng, L.: \u0026laquo;The MNIST Database of Handwritten Digit Images for Machine Learning Research [Best of the Web],\u0026raquo; \u003cem\u003eIEEE Signal Processing Magazine\u003c/em\u003e, cilt 29, pp. 141\u0026ndash;142, November (2012)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKramer, O.: \u0026laquo;Scikit-Learn,\u0026raquo; %1 i\u0026ccedil;inde \u003cem\u003eMachine Learning for Evolution Strategies\u003c/em\u003e, pp. 45\u0026ndash;53. Springer International Publishing (2016)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHodgkin ve, A.L., Huxley, A.F.: \u0026laquo;A quantitative description of membrane current and its application to conduction and excitation in nerve\u0026raquo;. Bulletin Math. Biology cilt. \u003cb\u003e52\u003c/b\u003e, 25\u0026ndash;71 (1952)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLapicque, L.M.: \u0026laquo;Recherches quantitatives sur l\u0026rsquo;excitation electrique des nerfs\u0026raquo;. J Physiol. Paris cilt. \u003cb\u003e9\u003c/b\u003e, 620\u0026ndash;635 (1907)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIzhikevich, E.M.: Dynamical Systems in Neuroscience: The Geometry of Excitability and Bursting. The MIT Press (2006)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eH. W. Kessels ve R. Malinow, \u0026laquo;Synaptic AMPA Receptor Plasticity and Behavior,\u0026raquo; \u003cem\u003eNeuron\u003c/em\u003e, cilt 61, pp. 340\u0026ndash;350, (2009)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMody, I., De Koninck, Y.: T. S. Otis ve I. Soltesz, \u0026laquo;Bridging the cleft at GABA synapses in the brain,\u0026raquo; \u003cem\u003eTrends in Neurosciences\u003c/em\u003e, cilt 17, pp. 517\u0026ndash;525, January (1994)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCover ve, T., Hart, P.: \u0026laquo;Nearest neighbor pattern classification,\u0026raquo; \u003cem\u003eIEEE Transactions on Information Theory\u003c/em\u003e, cilt 13, pp. 21\u0026ndash;27, January (1967)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCortes ve, C., Vapnik, V.: \u0026laquo;Support-vector networks,\u0026raquo; \u003cem\u003eMachine Learning\u003c/em\u003e, cilt 20, pp. 273\u0026ndash;297, September (1995)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCheeseman, P.C., Self, M., Kelly, J., Taylor, W.: D. Freeman ve J. C. Stutz, \u0026laquo;Bayesian Classification.,\u0026raquo; %1 i\u0026ccedil;inde \u003cem\u003eAAAI\u003c/em\u003e, (1988)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXanthopoulos, P., Pardalos, P.M., Trafalis, T.B., Xanthopoulos, P.: P. M. Pardalos ve T. B. Trafalis, \u0026laquo;Linear discriminant analysis,\u0026raquo; \u003cem\u003eRobust data mining\u003c/em\u003e, p. 27\u0026ndash;33, (2013)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRosenblatt, F.: \u0026laquo;The perceptron: A probabilistic model for information storage and organization in the brain\u0026raquo;. Psychological Rev. cilt. \u003cb\u003e65\u003c/b\u003e, 386\u0026ndash;408 (1958)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHines ve, M., Carnevale, T.: \u0026laquo;NEURON Simulation Environment,\u0026raquo; %1 i\u0026ccedil;inde \u003cem\u003eEncyclopedia of Computational Neuroscience\u003c/em\u003e, cilt 9, Springer New York, pp. 2012\u0026ndash;2017. (2015)\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"signal-image-and-video-processing","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"sivp","sideBox":"Learn more about [Signal, Image and Video Processing](http://link.springer.com/journal/11760)","snPcode":"11760","submissionUrl":"https://submission.nature.com/new-submission/11760/3","title":"Signal, Image and Video Processing","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"Spike pattern generation, spiking neural networks, population coding, feature encoding, supervised learning, biological neuron models, neuromorphic classification","lastPublishedDoi":"10.21203/rs.3.rs-6709572/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6709572/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThis study introduces SPIGEN, a configurable spike pattern generator designed to convert continuous-valued input features into fixed-length binary spike representations suitable for spiking neural networks. Based on a population coding framework, SPIGEN enables flexible configuration through multiple quantization and encoding schemes. Its primary goal is to transform continuous features into time-independent spike patterns that preserve inter-class separability and intra-class consistency. The study includes two experimental stages to evaluate the effectiveness of the proposed method. First, SPIGEN was used to generate spike patterns from two benchmark datasets\u0026mdash;IRIS and Digits\u0026mdash;and a synthetically generated Gaussian dataset. The resulting spike patterns were classified using conventional machine learning algorithms and a single-layer perceptron to examine how well the transformation process retained the information content of the original features. The results indicated that SPIGEN patterns yielded classification performances comparable to the original features, and the choice of encoding configuration significantly influenced the outcome. In the second stage, the generated spike patterns were applied as synaptic inputs to biologically inspired neuron models, where '1' indicates a spike and '0' denotes silence. When both excitatory and inhibitory synapses were used during training, the evaluation results suggested that SPIGEN is a viable solution for representing continuous input features in biologically plausible spiking neural networks.\u003c/p\u003e","manuscriptTitle":"SPIGEN: A Configurable Population-Coded Spike Pattern Generator for Single-Spike Learning in Spiking Neural Networks","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-06-02 10:25:24","doi":"10.21203/rs.3.rs-6709572/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-05-29T15:21:57+00:00","index":"","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-05-29T15:21:22+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-05-22T09:34:24+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-05-22T09:33:53+00:00","index":"","fulltext":""},{"type":"submitted","content":"Signal, Image and Video Processing","date":"2025-05-20T16:06:26+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"signal-image-and-video-processing","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"sivp","sideBox":"Learn more about [Signal, Image and Video Processing](http://link.springer.com/journal/11760)","snPcode":"11760","submissionUrl":"https://submission.nature.com/new-submission/11760/3","title":"Signal, Image and Video Processing","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"c40f7b3a-798a-41c0-9e3a-bbcf4fb6d3aa","owner":[],"postedDate":"June 2nd, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"in-revision","subjectAreas":[],"tags":[],"updatedAt":"2026-01-06T09:24:21+00:00","versionOfRecord":[],"versionCreatedAt":"2025-06-02 10:25:24","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6709572","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6709572","identity":"rs-6709572","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}