A scalable algorithm based on spike train distance to select stimulation patterns for sensory feedback

In this work, we develop an algorithm to optimally select a set of output spiking channels from a set of input sensing channels in order to deliver naturalistic sensory feedback to an amputee. The growing discrepancy between the number of sensing channels (103) and the number of stimulation channels (102) in neural interface devices leads to a combinatorially explosive search problem that quickly becomes intractable through brute force methods. This algorithm is tested using a dataset of 13 textures that was collected using a tactile sensor array with 9 tactile pixels (taxels). The sensor readings are transformed into neuron-like (neuromorphic) spiking activity that resemble the output of slowly adapting (SA) and rapidly adapting (RA) mechanoreceptors in the skin which encode different features of tactile information. The channel selection algorithm uses a spike train distance metric to create an optimal set of output channels that are highly differentiated and complementary. The effectiveness of the channel selection algorithm is tested by comparing texture classification performance between the chosen set and a set of random channels. The selected channels have slightly better classification accuracy but a much smaller standard deviation. As the input channels get noisier and grow in number, the channel selection algorithm will provide even better accuracy in comparison to random selection and much lower computational cost in comparison to brute force search. This work paves the way to provide sensory feedback systems that will deliver more informative stimulation patterns to amputees.