Design and implementation of a Convolution event-based network with Offline Learning

Nowadays, Convolutional Neural Networks (CNNs) are exploited to solve different tasks, but their increasing complexity means an increase in the power consumption of these architectures which limits their applications. The introduction of Spiking Neural Networks (SNNs) is an important step to overcome this limit. They work in the same way as the behaviour of our brain and they are organized in layers of biological neurons, which receives spikes, elaborate them and solve the task. Spikes mean a reduction of the complexity of the operation, due to the substitution of the Multiply and Ac-cumulate operation with a simple Select and Accumulate, which is traduced into a reduction of the computational power. The work is focused on the implementation VHDL of a convolution event-based neural network with offline learning based on a script PyTorch, used to recognize handwritten digits based on MNIST dataset. The architecture is organized with a convolutional layer, which extracts the features of the input, a max pooling layer, which reduces the noise and the image dimen-sion, and two layers of Izhikevich Neuron used to classify the digits from 0 to 9. To train the architecture, a PyTorch script has been written. To describe an event-drive convolutional neural network, PyTorch is extended with Norse library. First, the architecture is tested with this software script, in order to train it and to find weights of the fully connected layer and kernel of the convolutional one and then, with the obtained values, the hardware has been tested in order to verify the correctness of the described architecture.