Method of Neural Network Training with Integer Weights

This paper presents the method of training the feedforward neuronal network with integer weights and integer input and output signals. Neural networks of this type are better suited for embedded systems and hardware implementation than neural networks with floating weights. Also, the new method allows training neural networks on embedded devices with limited resources. This algorithm has been developed taking into account that the resulting integer weights require less memory for storage, and arithmetic operations over them are easier for hardware implementation. A simple activation function is proposed that works with integer data and limits weights and bias in a certain range during neural network training. The simulation results show that the new training method gives a similar result while training with floating-point arithmetic for XOR operation. To validate the robustness of the new algorithm to train neural networks with integer arithmetic, a prototype was created based on the CYAT817 microcontroller with the Cortex-M0+ core.