Spiking Sparse Coding Algorithm with Reduced Inhibitory Feedback Weights.

In this paper we demonstrate that a sparse coding algorithm using spiking neurons can be designed to have reduced inhibitory feedback connections by modifying SAILNet [1]. This modification makes it easier for a circuit designer to implement sparse coding in hardware because there is less circuitry for learning and fewer parameters to store in memory. We show our modification and compare it with SAILNet. Our analysis shows that the modification is more sparsely active than the original algorithm without significantly affecting the reconstruction of the input stimulus. It is also experimentally shown that by tuning the value of inhibitory synapse strength, sparsity can be controlled. The learned receptive fields are similar to those of SAILNet and retain information from stimulus such that reconstructed stimulus can be identified by a convolutional neural network for classification.