From beasts to bytes: Revolutionizing zoological research with artificial intelligence.

The field of paleozoology, dedicated to the study of prehistoric life through fossil analysis, has traditionally demanded significant labor and time. In recent years, however, advancements in AI algorithms capable of learning, prediction, and automation have revolutionized this field by enhancing fossil analysis, species identification, and the extrapolation of functions and behaviors from fossil evidence ( Figure 3G ). AI has been effectively applied for the classification of animal fossils based on both text and image data. For instance, in their study on marine invertebrates, Kopperud et al. ( 2019 ) used an automated pipeline based on supervised machine learning to extract observations of fossils and their inferred ages from unstructured text, achieving high accuracy compared to human annotators and greatly reducing the time required. In a broader study, Hoyal Cuthill et al. ( 2020 ) used a machine learning spatial embedding method to measure distance among 171 231 species and analyze the impacts of speciation and extinction based on an evolutionary decay clock. In regard to primates, AI has been utilized to infer behaviors from fossils based on species classification ( Marcé-Nogué et al., 2020 ; Püschel et al., 2018 ). Püschel et al. ( 2018 ) applied machine learning classification methods to differentiate between various locomotor categories in primates, using both biomechanical and morphometric data. Similarly, Marcé-Nogué et al. ( 2020 ) used an SVM algorithm to classify primate fossils into broad ingesta-related hardness categories. The application of deep learning algorithms for image data analysis has become increasingly prevalent in resolving classification issues related to animal fossils. CNNs have demonstrated considerable accuracy in the classification of modern and ancient bone surface modifications within multiple mammals ( Domínguez-Rodrigo et al., 2020 ). Using both unsupervised and supervised AI algorithms, studies have achieved high accuracy (88%–98%) in classifying carnivore species based on 3D modeling and geometric morphometrics of tooth pits ( Courtenay et al., 2021 ). Furthermore, deep CNNs have distinguished between theropod and ornithischian dinosaur tracks based on outline silhouettes, consistently demonstrating superior performance compared to human experts ( Lallensack et al., 2022 ).

Deep learning models, particularly CNNs, are favored by researchers as predictive models due to their local receptive fields and convolutional architecture, which make them effective at processing image data. However, their efficacy in handling sequential data, such as text and video recordings, is less satisfactory compared to linguistic models such as LSTM and BERT. The main limitation of CNNs is their inability to effectively capture and preserve the positional dependencies inherent in sequential data. Thus, zoological researchers may adopt different trajectories, which is not necessarily problematic but warrants further exploration. Supervised learning is also prevalent in zoological studies, commonly chosen for fitting models and performing classification tasks. In contrast, unsupervised learning is less frequently used, primarily for generative and clustering tasks, due to the absence of supervised labels. Reinforcement learning also draws inspiration from zoological studies exploring animal decision-making behavior derived from neuronal activation. Such integration is expected, as the field of zoological research emphasizes using AI models to address specific, domain-related challenges spanning supervised, unsupervised, and reinforcement learning. The integration of AI into zoological research promises a paradigm shift, enhancing efficiency, accuracy, and scope ( Wang et al., 2023a ). AI will enhance data collection and interpretation by increasing efficiency, reducing time expenditure, and improving accuracy, especially in handling complex data like intricate sounds and high-resolution images ( Shen et al., 2022b ). AI will enable real-time, non-invasive monitoring of wildlife, in contrast to traditional methods that are often invasive and limited. AI technologies also promote objectivity in behavioral analysis by automating processes and providing quantitative metrics. The focus of zoological research is shifting from mere descriptive to predictive and explanatory models. This technology will also improve the verifiability and reproducibility of research, influencing the broader acceptance of theories. The interdisciplinary nature of AI also allows for the integration of computer science, statistics, and biology, facilitating the development of new theoretical frameworks. AI further democratizes zoological research by enabling public participation in data collection, such as through social media-based wildlife monitoring ( Foglio et al., 2019 ). Paradigm shifts are not usually completed in the short term but are gradual, evolving transformations. However, as the application of AI in zoology becomes increasingly widespread, we anticipate a series of profound changes in this field, including in the evolution of theoretical frameworks. These changes will advance our understanding of the complexity of the animal world in a more comprehensive and in-depth manner. In the future, beyond the establishment and use of cameras in nature reserves, AI will also be able to utilize satellite images, climate data, and drones to detect changes in animal populations or potential threats like poaching, helping to predict areas where endangered species can reproduce and thrive. The integration of AI with biological recording devices will also revolutionize the study of animal migration, behavior, and physiology under natural conditions. The potential development of a global platform will enable zoologists worldwide to collaborate, share data, and utilize AI tools collectively for analysis. Zoological research can also directly influence model design, drawing insights from animal neuroscience and behavioral studies ( Yu, 2016 ). Furthermore, the zoological domain provides a vast and diverse dataset, supporting comparative analysis of models across different architectures and theoretical paradigms. As a result, it serves as an ideal experimental platform for AI models requiring extensive data resources. Therefore, the relationship is not just about applying AI to zoology but is mutually beneficial for both disciplines.

AI was first defined by Stanford Professor John McCarthy in 1955 as a “the science and engineering of making intelligent machines” ( Shabbir & Anwer, 2018 ). Machine learning, a fundamental branch of AI, is characterized by the capability of systems to autonomously learn from large datasets ( de Souza Filho et al., 2020 ). Machine learning attempts to utilize experience, usually in the form of data, to improve model performance and the process of learning ( Sarker, 2021 ). Therefore, machine learning research primarily focuses on the development of algorithms that generate models from data, termed “learning algorithms”. In the field of zoology, machine learning is helping to shed light on tasks such as species classification, behavior identification, animal population size prediction, bird sound recognition, and nonhuman animal language learning ( Layton et al., 2021 ; Norouzzadeh et al., 2018 ). These specific tasks are broadly categorized into three types of machine learning, differentiated by their respective data training approaches: supervised, unsupervised, and reinforcement learning ( Dönmez, 2013 ) ( Figure 1C ). Supervised learning, one of the most widely used machine learning methods, relies on explicit datasets labeled by experts ( Jiang et al., 2020 ). Supervised learning algorithms build models to identify relationships within a set of feature-label pairs, utilizing the label, also known as the target, for training the system ( Li, 2017c ). These algorithms fall into two main categories: classification (discrete modeling) and regression (continuous modeling). Both categories are predictive modeling techniques, differing only in their target (response) variables. In classification, the target variable is discrete and takes the form of categories (class labels). For example, animal species identification ( Binta Islam et al., 2023 ), i.e., species classification, relies on learning from diverse data types, such as images, footprints, sounds, and videos, collected from various animals. More importantly, these data are manually annotated with labels indicating the species to which each sample belongs, with the labels corresponding to a predefined list of species. In this task, a model is trained to predict the species name when presented with new input data, such as animal photos. In contrast, in regression tasks, the target variable is continuous rather than discrete. For example, predicting the weight of cultivated cattle constitutes a continuous regression analysis, as the target variable, weight, is continuous. Unsupervised learning ( Li, 2017a ) analyzes and clusters unlabeled datasets. Unlike in supervised learning, the inputs in this approach are usually raw data without available labels. These algorithms uncover hidden patterns in data without requiring human intervention, thus termed “unsupervised”. Unsupervised learning models are used for three main tasks: clustering algorithms, which aim to discover unknown subgroups in unlabeled data based on their similarities or differences; dimensionality reduction techniques, which aim to minimize the dimensionality of data by discarding redundant or non-task-relevant information; and anomaly detection, which aims to identify observations that may have originated from different data generation processes. For example, application of the clustering task has enabled researchers to uncover the social structures within jackdaw populations by analyzing unlabeled data of visitation times of each individual ( Valletta et al., 2017 ). Reinforcement learning ( Li, 2017c ) involves a family of algorithms that typically operate sequentially. These algorithms are trained through interactions between agents and the (virtual) environment and applied to tasks where learning depends on executed actions and resulting consequences. A notable success is the AlphaGo computer program ( Wang et al., 2016 ), which outperformed human players based on its integration of deep neural networks and reinforcement learning techniques. Frankenhuis et al. ( 2019 ) have advocated for the broader application of reinforcement learning methods in behavioral ecology to address the challenge of inferring the unknown reward functions of agents and to explore how biological mechanisms tackle developmental and learning problems. In addition, AI tasks can also be classified based on the type of input data they process. The three typical subfields include computer vision (CV), natural language processing (NLP), and multi-modal learning. In the field of CV, the primary types of input data include images (e.g., grayscale, color, and binary images) and three-dimensional (3D) representations (e.g., point clouds ( Liao et al., 2021 )). In the field of NLP, the main data types include text data, speech data (e.g., voice and sound recordings), and time-series data (e.g., motion trajectories captured over time). In contrast, multi-modal learning ( Lahat et al., 2015 ) integrates information from various data sources, ranging from CV to NLP and including images, text, and voice recordings. This integrative approach affords a richer data representation and more effectively captures relationships, thereby enabling a more complex and comprehensive analytical understanding. CV seeks to automate tasks that the human visual system can perform and is concerned with the automatic extraction, analysis, and understanding of useful information from a single image or a sequence of images (e.g., videos). CV involves the development of a theoretical and algorithmic basis to achieve automatic visual understanding ( Farahbakhsh et al., 2020 ). In this field, typical tasks ( Chai et al., 2021 ) include image classification, retrieval, object detection, semantic segmentation, instance segmentation, object localization, action recognition, and object tracking, among other more specific tasks. Image classification ( Lorente et al., 2021 ) is a fundamental task in CV that aims to categorize an image as a whole under a specific label. When the classification becomes highly detailed or reaches instance-level, it is often referred to as image retrieval ( Chen et al., 2023b ), which also involves the identification of similar images within a large database. Object detection ( Zou et al., 2023 ) aims to detect and locate objects of interest within an image or video. Object localization focuses on locating an instance of a particular object category in an image, typically by specifying a tightly cropped bounding box centered on the instance. In the literature, “object localization” refers to locating one instance of an object category, whereas “object detection” focuses on locating all instances of a category in a given image ( Chai et al., 2021 ). Semantic segmentation aims to categorize each pixel in an image into a class or object, producing a dense pixel-wise segmentation map where each pixel is assigned to a specific class or object. Instance segmentation involves identifying and separating individual objects within an image, including detecting the boundaries of each distinct object of interest and assigning a unique label to each one. CV is skilled at handling various scenes related to images and videos. NLP enables a computer system to automatically process and analyze sequence data, even extracting latent information by understanding the syntax in sequences, which is natural for humans ( Young et al., 2018 ). Notably, NLP can capture dense vector representations, known as feature embeddings, from raw sequence data ( Pilehvar & Camacho-Collados, 2021 ). Based on these feature representations, various language models have been created for downstream NLP tasks ( Young et al., 2018 ). Within the context of zoological research, several closely related NLP tasks are commonly applied, including tokenization ( Sodhar et al., 2020 ) for text data processing, text classification for video recognition ( Xu et al., 2016 ), and language modeling for text or gene sequences. With advancements in deep learning models, NLP has the potential to revolutionize research in the zoological domain. This subsection provides an overview of deep learning from various perspectives, including main concepts, architectures, and computational tools. The aim is to highlight the most important aspects of deep learning and serve as an instructive guideline for zoologists seeking to utilize this tool. Due to its extraordinary learning capabilities, deep learning technology, which originated from artificial neural networks, has become a hot topic in the context of AI, with wide application in various areas such as CV, NLP, and speech recognition. Notably, between 2018 and 2021, the number of publications on neural network algorithms has shown a five-fold increase ( Shine & Murphy, 2022 ). AI models are broadly categorized into shallow or deep learning models based on the number of linear or non-linear transformations the input data undergo before yielding an output ( Ahmad et al., 2018 ). Shallow models typically convert inputs once or twice before transmitting outputs, while deep models, derived from conventional neural networks, commonly convert inputs multiple times ( Meir et al., 2023 ). As a result, deep models can learn more complex patterns, thereby facilitating end-to-end learning without the need for manual feature engineering and exhibit robust performance in CV and sequential data analysis tasks. The introduction of backpropagation (BP) algorithms for artificial neural networks (commonly referred to as neural networks) in the 1980s ushered in an era of machine learning dominated by statistical models ( Janiesch et al., 2021 ), which continues to this day. In the 1990s, challenges such as overfitting and slow training speed in artificial neural networks led to the proposition of various other shallow machine learning models, including support vector machines (SVM), boosting techniques, and maximum entropy methods (e.g., logistic regression) ( Xu et al., 2021 ). SVM and boosting are examples of hidden nodes rather than models (e.g., logistic regression) that utilize hidden nodes. Geoffrey Hinton, along with Yoshua Bengio and Yann LeCun, has been a persistent advocate for the advancement of neural networks, playing a significant role in the development of a practical and feasible deep learning framework ( Wang & Duan, 2021 ). The effectiveness of machine learning algorithms is highly dependent on the integrity of the input data representation. Therefore, feature engineering has long been an important research area in machine learning, aiming to extract features from raw data with considerable human investment. In contrast, deep learning algorithms automate the feature extraction process, thus reducing reliance on extensive human labor and domain expertise to extract salient features. These algorithms possess a multi-layer data representation architecture, with the first layer extracting low-level features and the last layer extracting high-level features ( Wang & Duan, 2021 ). Due to its considerable success, deep learning has emerged as a prominent research trend. In this context, convolutional neural networks (CNNs), recurrent neural networks (RNNs), and pretrained foundation models (PFMs) have been increasingly employed in zoological research. Within the field of deep learning, CNNs are the most commonly used algorithms, with extensive application in image recognition, speech recognition, and NLP. The three principal advantages of CNN are equivariant representation, sparse interaction, and parameter sharing. Unlike traditional fully connected networks, CNNs leverage shared weights and localized connections to fully exploit two-dimensional (2D) input data structures, such as those found in image data. This operation employs a very small number of parameters, which simplifies the training process and accelerates network speed. This concept mirrors the functioning of cells in the visual cortex, as elucidated by Alzubaidi et al. ( 2021 ), where each unit is responsive to only a subset of the visual field, thereby capturing spatial localities within the input similar to the application of localized filters. A common type of CNN, akin to a multi-layer perceptron, features numerous convolutional layers preceding the subsampling (pooling) layer and concluding with a fully connected layer ( Figure 2 ). Concise architectures of three deep learning models Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), and Pretrained Foundation Models (PFM). The high performance achieved by CNN architectures in challenging benchmark task competitions indicates that innovative architectural concepts and parameter optimization can improve CNN performance in various visually related tasks ( Khan et al., 2020 ). The exploration of grid topology data (image and time-series data) by LeCun et al. ( 1989 ) marked the initial recognition of the capabilities of CNNs. Since 2012, different innovations have been proposed regarding CNN architecture. The performance improvements in CNNs can primarily be attributed to the reconstruction of processing units and design of new modules. With the introduction of AlexNet ( Krizhevsky et al., 2017 ) and its exceptional performance on ImageNet datasets, the application of CNNs has become increasingly popular. The development of the Inception module by the Google team, characterized by its split-transform-merge strategy, marked a substantial advancement in CNN architectures. Its novel introduction of intra-layer branching facilitated feature extraction across varying spatial dimensions. In 2015, ResNet revolutionized CNN training by introducing residual connections ( He et al., 2015 ), a concept incorporated in many subsequent networks, including Inception ResNet, wide residual networks, and ResNext. Similarly, certain architectures, including wide residual networks, Pyramid Nets, and Xception, have introduced multi-layer transformations, implemented through additional cardinality and increased width. Typical deep CNN models, in addition to a fully connected layer (sometimes not included because of Global Average Pooling), also include convolutional and pooling layers, used to extract meaningful features from locally associated data sources and reduce the number of parameters, respectively ( Khan et al., 2020 ) ( Figure 2 ). Compared to other methods, the minimal preprocessing requirements of CNNs have solidified their status as the de-facto standard computation framework in CV. The achievements of CNNs have garnered widespread attention, both inside and outside academia, leading to the proliferation of diverse CNN models. For instance, the partialS/HIC model ( Xue et al., 2021 ), which is based on a CNN for image processing, addresses the increasing demand for scanning tools capable of tracking in-progress evolutionary dynamics. Similarly, the DeepBehavior toolbox ( Arac et al., 2019 ), which integrates many different CNNs, can automatically analyze animal behavior from both video and 3D image data. The landscape of CNN-based frameworks includes ResNet, MobileNet, DenseNet, ShuffleNet, EfficientNet, R-CNN, and YOLO ( Zhu & Zhang, 2018 ). Among them, YOLO is an advanced algorithm for object detection, and includes different versions such as YOLOv2, YOLOv3, YOLOv4 ( Bochkovskiy et al., 2020 ), and YOLOR ( Wang et al., 2021 ). Each algorithm designed for object detection has the capability to identify objects both in real-time and with high accuracy. RNNs, designed to handle sequential data (e.g., text, speech, and time series data), are commonly employed in the field of deep learning ( Figure 2 ), including speech processing and NLP ( Lipton et al., 2015 ). RNNs learn the features of time series data by memorizing previous inputs in the internal state of NNs. They can also predict future information based on past and present data but struggle to learn long sequences structures due to the vanishing or exploding gradient issue. Long short-term memory (LSTM) ( Staudemeyer & Rothstein Morris, 2019 ) networks and their variants, such as the gated recurrent unit (GRU) networks ( Chung et al., 2014 ), have resolved the gradient issue using various gates that control how information flows. These algorithms can be used in fields requiring the analysis of sequential data and the prediction of future events based on present data. Research in NLP frequently addresses time series data, analogous to that encountered in zoology, such as text, sound recordings, and sequence data. To capture the positional dependencies inherent in sequential data and retain abundance information from raw data, several deep learning-based models have been developed. RNN-based models (like RNN ( Rumelhart et al., 1986 ), LSTM ( Hochreiter & Schmidhuber, 1997 )) and transformer-based models ( Vaswani et al., 2017 ) (like BERT) have proven effective in automatically modeling sequential data and conducting downstream task prediction and analysis. Transformer-based models are fundamentally structured around the attention mechanism and its derived framework, multi-head attention ( Choi & Lee, 2023 ). Leveraging this core architecture, transformer-based models can overcome the limitations of RNN models, which cannot parallelize input processing across all time steps, while still effectively capturing the positional dependencies inherent in sequential data in tasks such as language translation, text summarization, image captioning, and speech recognition. The transformer includes encoder and decoder structures for processing sequence inputs and generating corresponding outputs, respectively ( Choi & Lee, 2023 ). The architectural variant BERT, which only incorporates the encoder structure, employs random sequence masking during its pre-training tasks, leading to superior outcomes in protein 3D structure prediction ( Lin et al., 2022b ) and single-cell annotation ( Yang et al., 2022 ). Pretrained foundation models ( Zhou et al., 2023 ) are essential and significant components of AI in the era of big data. These models demonstrate enhanced proficiency in multi-task learning with large-scale datasets and exhibit increased efficiency during fine-tuning for targeted, smaller-scale tasks, resulting in rapid data handling capabilities ( Bommasani et al., 2021 ) ( Figure 2 ). The most famous application among them is ChatGPT, a conversational model derived from the generative pre-trained transformer architecture developed by OpenAI ( Gozalo-Brizuela & Garrido-Merchan, 2023 ). ChatGPT applies reinforcement learning from human feedback (RLHF) ( Bai et al., 2022 ), a promising approach for aligning large language models with human intent, i.e., pretrained language models ( Wulff et al., 2023 ). Many open-source pretrained language models are currently available, operable on individual computing systems and trainable on private datasets, including Llama, Alpaca, Vicuna, and Falcon models ( Zhang et al., 2023 ). Given their success ( Wang et al., 2023b ) in various general-domain NLP tasks, these open-source large language models exhibit significant potential for application when fine-tuned using knowledge-based instruction data. Inspired by the success of pretrained language models in NLP, pre-trained visual models in the field of CV have also achieved great success. These models are pre-trained on massive image datasets and can analyze image content and extract rich semantic information. Furthermore, multi-modal visual models like CLIP ( Radford et al., 2021 ) and ALIGN ( Cohen, 1997 ; Lahat et al., 2015 ) use contrastive learning to align textual and visual information. This alignment allows the pre-trained models to apply learned semantic information to the visual domain, thereby facilitating efficient generalization in downstream tasks, including zoological applications. Generative models and contrastive learning are two other important types of models. Generative models gained popularity after the introduction of generative adversarial networks (GANs) in 2014, which formed the foundation for many subsequent architectures, including CycleGAN ( Zhu et al., 2017 ), StyleGAN ( Karras et al., 2019 ), and DiscoGAN ( Kim et al., 2017 ). Unlike generative models, contrastive learning is a discriminative approach that aims to group similar samples closer together and diverse samples farther apart ( Jaiswal et al., 2021 ). To achieve this, a similarity metric is used to measure the closeness of two embeddings. Self-supervised learning, a type of unsupervised learning ( Wang, 2022 ), integrates both generative and contrastive approaches. Notably, it utilizes unlabeled data to learn underlying representations, thereby avoiding the labor-intensive task of data labeling. Thus, this approach offers the potential for better utilization of unlabeled data in zoological research. Given that deep learning has proven more effective in data extraction, feature representation, and prediction than non-deep learning models, our emphasis is on discussing deep learning models used in zoological research and explaining why they are more effective compared to non-deep learning models, especially when dealing with unstructured data such as images, text, videos, and sequence data. Furthermore, while supervised task is frequently discussed in this review due to their relevance to species classification and identification (fundamental concerns in Zoology), we also address unsupervised and reinforcement learning. A variety of datasets are available for research in zoology, each tailored to specific tasks, as summarized in Supplementary Table S2. These datasets primarily encompass text and image data. Text data can be processed using a range of models, such as RNN, transformer, and pretrained foundation models, depending on the specific task requirements. Image data, including videos, are compatible with models developed for CV tasks. Among these datasets, the Paleobiology Database, maintained by an international consortium of non-governmental paleontologists, is a publicly accessible repository of paleontological data ( Alroy et al., 2008 ). The AP-10K dataset ( Yu et al., 2021 ) represents the first comprehensive resource for general animal pose estimation, featuring 10 015 images from 23 animal families and 54 species, with high-quality keypoint annotations. This highly versatile dataset is suitable for supervised, self-supervised, semi-supervised, and cross-domain transfer learning, as well as intra- and inter-family domain analyses, with annotation files provided in Common Objects in Context (COCO) format. The KaoKore dataset, established by the ROIS-DS Center for Open Data in the Humanities, comprises a curated collection of facial expressions and has been publicly accessible since 2018 ( Tian et al., 2020 ). Complex unstructured data in zoological research, including images, videos, sounds, and text, present considerable challenges for AI application. Addressing these challenges often requires expert selection of appropriate models and significant data preprocessing efforts. The following sections provide detailed discussion on AI models tailored for different data types and tasks (Supplementary Table S3), as well as their specific processing procedures. This information should help researchers in selecting models best suited for their specific data and research goals.

As discussed in the introduction, there is a lag in the adoption of AI in zoological research ( Figure 1A , B). One reason for this delay is the unfamiliarity of zoologists with various models, coupled with the challenges arising from data format complexity, data insufficiency, and reliance on small sample learning tasks. Comprehensive zoological research includes complex unstructured data, spanning images, videos, sound recordings, text sequences, and protein structures, presenting significant challenges for AI model applications, requiring experts to select appropriate models for specific data types and, in some instances, expend additional effort in data processing. Small sample learning is another common challenge, especially when studying specific species or ecosystems, with the extreme rarity of certain animal species posing considerable challenges in gathering adequate data. Furthermore, while identifying animals based on their sounds can provide valuable insights, gathering high-quality sound data can be time and resource-intensive, leading to limitations in sample sizes ( Bravo Sanchez et al., 2021 ). In addition, observing animal behavior typically requires time and effort, often resulting in limited data and constraints on in-depth studies of specific behaviors ( Arac et al., 2019 ). These circumstances give rise to imbalanced datasets and fewer samples from rare species for training neural networks ( Høye et al., 2021 ). Although the application of AI models in zoology has been slow compared to the broader biological sciences, the introduction of new technologies in specialized subfields often comes with an inherent time lag. To address these challenges, it is recommended to enhance data collection initiatives, focusing on long-term accumulation and integration with existing databases to mitigate data insufficiency. In situations of limited sample sizes, employing pre-training and transfer learning methods would be beneficial (( Høye et al., 2021 ; Vélez et al., 2022 ). For data concerning rare species or of low quality, applying data augmentation methods such as image manipulation and geometric transformations can expand training datasets ( Klasen et al., 2022 ). Model-agnostic meta-learning ( Shui et al., 2023 ) and multiset feature learning ( Jing et al., 2021 ) could offer innovative solutions to these issues ( Shui et al., 2023 ; Jing et al., 2021 ). When encountering species not present in existing databases, the challenge becomes even more significant. Addressing this situation necessitates the application of multiple class anomaly/novelty detection or open set/world recognition ( Perera & Patel, 2019 ; Turkoz et al., 2020 ). Moreover, simpler models such as logistic regression, K-Nearest Neighbor Regression, and SVM may be more suitable in these cases. To address generalization issues, emphasis should be placed on data integrity through meticulous cleaning and quality control processes. Leveraging pre-trained models and fine-tuning them on specific tasks can also aid in adapting to varied species data ( Lin et al., 2023 ). Lastly, the broader dissemination and implementation of these AI technologies should be encouraged, alongside fostering collaborative efforts between AI experts and domain specialists. Supervised learning plays a pivotal role in zoological research due to its close association with species classification, behavior identification, and regression tasks in feeding behaviors. Supervised learning requires high-quality annotated data to train models and produce highly accurate predictions. As discussed in the “Animal Classification and Resource Protection” section, certain species classification models have shown suboptimal accuracy ( Bravo Sanchez et al., 2021 ), which may be partly attributed to a lack of labeled data for supervised learning. To address this issue, techniques like semi-supervised, which combines supervised and unsupervised learning, and weakly supervised learning have been applied in the field of AI, offering potential methodologies for zoological research.

This review comprehensively examines the integration of AI within zoological research, outlining the primary tasks, core models, datasets, and the challenges encountered in this field. AI algorithms have been instrumental in advancing various subfields, managing complex data forms, and achieving leading-edge predictive accuracy. The combination of AI and zoological research harnesses the computational power of “bytes” to successfully decipher the inner secrets of “beasts”. Given the increasing demand for analyzing data generated from laboratories, natural habitats, and agricultural settings—including but not limited to images, text, tables, videos, gene sequences, and molecular structures—the strategic selection and application of AI models tailored to specific scientific inquiries is expected to gain momentum.

Artificial intelligence (AI) stands at the forefront of modern scientific innovation ( Wang et al., 2023a ). Its research applications range from medical diagnostics ( Moor et al., 2023 ) to climate tracking ( Gore, 2022 ), with the scope of AI research expanding continuously. Of note, AI is making particularly significant strides in zoological research ( Romero-Ferrero et al., 2019 ). One challenge of zoological research, a discipline focused on animal classification, behavior, physiology, development, genetics and evolution, disease modeling, and paleozoology, is the management and interpretation of extensive and complex datasets. The rapid emergence of advanced AI techniques, such as machine learning ( Jordan & Mitchell, 2015 ) and, in particular, deep learning ( Hinton et al., 2006 ), as well as the emergence of big data, has marked the beginning of an era of intelligent data-centric zoological research. Although AI has been popular for some time, its incorporation into zoological research has not kept pace with its application in other biological fields ( Figure 1A , B; Supplementary Table S1). Thus, the question arises as to why AI technologies have not been promptly adopted in animal research. A possible factor for the lack of application may be the inefficiency of computational resources and scarcity of expansive zoological datasets. Additionally, zoologists may lack the foundational knowledge required to understand and implement these approaches, creating uncertainty regarding the selection of models suitable for their objectives. Moreover, the rapid and continuous evolution of complex AI model architectures, like Bidirectional Encoder Representations from Transformers (BERT) ( Devlin et al., 2019 ), make it challenging for zoological researchers to stay current. As access to advanced computational tools and comprehensive zoological datasets expands, it may pave the way for broader adoption of these algorithms in mainstream research. Nevertheless, unfamiliarity with these techniques persists among many zoologists, necessitating a foundational understanding of when, why, and how to employ these methods, as well as what type of data is suitable for their application. Branches and applications of AI A: Timeline of various AI models and their applications in biological and zoological research. B: Word cloud charts displaying counts (Supplementary Table S1) of different AI models used in biological and zoological research, represented after log 2 transformation. C: Branches of AI, addressed tasks, and representative models. AdaBoost: Adaptive Boosting; A3C: Asynchronous Advantage Actor-Critic; BERT: Bidirectional Encoder Representations from Transformers; CNN: Convolutional Neural Network; CV: Computer Vision; DDPG: Deep Deterministic Policy Gradient; DQN: Deep Q-Network; EM: Expectation Maximization; GAN: Generative Adversarial Network; GNN: Graph Neural Network; KNN: K-Nearest Neighbors; LDA: Latent Dirichlet Allocation; LSTM: Long Short-Term Memory Network; MLP: Multi-Layer Perception; NLP: Natural Language Processing; NN: Neural Network; PCA: Principal Component Analysis; PPO: Proximal Policy Optimization; ResNet: Residual Neural Network; RF: Random Forest; RNN: Recurrent Neural Network; SOM: Self-Organizing Map; SVM: Support Vector Machine; t-SNE: t-distributed Stochastic Neighbor Embedding; XGBoost: Extreme Gradient Boosting. In this review, we present an introduction to AI and its primary tasks, elucidating the key models, datasets, and challenges faced. We also explore the intersection where beasts meet bytes, examining how AI applications are revolutionizing diverse areas of zoological research. By analyzing real-world case studies and predicting future directions, we offer a comprehensive overview of the role of AI in deepening our understanding of the animal kingdom and the potential fields it may unlock in the coming years.

Supplementary data to this article can be found online. Click here for additional data file.