Figure 3.

(A) Illustration of HB-STDP based unsupervised training methodology for $kerne{l}_{ij}^l$ connecting the i^th input spike map in layer “l−1” ($ma{p}_i^{l-1}$) to the j^th output spike map in layer “l” ($ma{p}_j^l$). The $kerne{l}_{ij}^l$ is updated using HB-STDP based on the spike timing difference between the spiking post-neuron in output $ma{p}_j^l$ and the corresponding pre-neurons in input $ma{p}_i^{l-1}$. The HB-STDP based weight updates are carried out on all the kernels in layer “l” based on the respective input and output maps. (B) Illustration of HB-STDP based mini-batch training methodology for mini-batch size of 2. The $kerne{l}_{ij}^l$ is now updated based on the average spike timing difference between the spiking post-neurons in the output mini-batch ($ma{p}_{j1}^l$ and $ma{p}_{j2}^l$) and the respective pre-neurons in the input mini-batch ($ma{p}_{i1}^{l-1}$ and $ma{p}_{i2}^{l-1}$).