Fig. 3: Accuracy (acc) and loss (mean squared error, mse) over epochs.
From: The backpropagation algorithm implemented on spiking neuromorphic hardware
The plot shows the mean over three training runs with random weight initialization and random dataset shuffles during training. Individual runs are shown with dots connected by a finer line. In addition to the accuracy and loss calculated using the last layer’s spikes, the top-1 accuracy is calculated from the last layer’s membrane potential. The accuracy calculation was performed off-chip using checkpoints stored from the on-chip trained network after each half training epoch. Note separate axis scaling for accuracy (left) and loss (right).
