Fig. 4: Conversion followed by fine-tuning for a VGG16 network.
From: High-performance deep spiking neural networks with 0.3 spikes per neuron
A VGG16 SNN network (right) is initialized (curly blue arrow) with the weights and biases (red) of a pre-trained VGG16 ReLU network (left), and then fine-tuned (FT) with gradient descent using the framework of the B1-model. The pre-trained ReLU network also undergoes further fine-tuning using standard backpropagation optimization.
