Deep neural network based on F-neurons and its learning

Abstract Artificial neural networks are widely used in data processing and Data Mining. In contrast to traditional artificial neural networks, deep neural networks employ many artificial neuron blocks and more than three layers. An increase in the number of neuron blocks and layers improves the approximation capabilities but leads to the effects of exploding and vanishing gradients. Deep neural networks often employ piece-wise activation functions like ReLU to overcome the effects of exploding and vanishing gradients. We propose F-neuron as an adaptive alternative to piece-wise activation functions. Like ReLU, F-neuron does not suffer from the effects of exploding and vanishing gradient. F-neuron changes its form during the training, can approximate any currently used function, and synthesize new task-specific functions. We show that F-neuron can synthesize new task-specific functions and achieve higher approximation quality in existing neural network architectures. We evaluate the performance of the F-neuron on two image classification datasets (Fashion-MNIST and CIFAR-10) in two different architectures (LeNet-5 and KerasNet).