Figure 1.

Illustration of ReStoCNet consisting of an input layer followed by stacked convolutional layers with Leaky-Integrate-and-Fire (LIF) spiking non-linearity, which are interconnected via binary kernels. The deeper convolutional layers receive residual inputs that are summed up with direct inputs from the preceding convolutional layer as depicted in the inset. The binary kernels forming the convolutional layers are trained using probabilistic Hybrid-STDP (HB-STDP) based layer-wise unsupervised training methodology. After all the convolutional layers are trained, the respective spike maps are spatially pooled using average pooling with 2 × 2 unit-weight kernels followed by Integrate-and-Fire (IF) spiking non-linearity to produce the pooled spike maps. The spike trains of the pooling layers are low-pass filtered to obtain their spiking activations over the time period for which the input is presented, which are fed to the fully-connected layer, trained using error backpropagation, for inference.