Source of accuracy increase after conversion to Akida spiking model

Determine the cause of the approximately 1% test-accuracy improvement observed when converting the quantized CNN-32-32-64 model to a BrainChip Akida AKD1000 spiking model using the MetaTF pipeline, and identify which specific aspects of the Akida conversion and execution (including quantization, neuron and synapse implementation, and inference timing) lead to accuracy changes relative to the quantized ANN baseline.

Background

In the hardware evaluation, Keras models were trained and then quantized using the BrainChip MetaTF SDK before conversion to spiking models for execution on the Akida AKD1000 neuromorphic processor. The quantized models typically experienced an accuracy drop relative to the original Keras models, which was largely recovered through retraining.

For the largest model tested (CNN-32-32-64), the Akida spiking model achieved about a 1% increase in accuracy compared to the retrained quantized model it was derived from. The paper notes that the reason for this improvement is not currently understood, highlighting uncertainty about the effects introduced by the conversion and execution on Akida hardware.