Neural Prostheses

Neural Decoding

Signal Classification

Spike bursting from a multi-channel array in the cortex. We are currently devising algorithms for decoding ENG recorded from peripheral nerve.
Typically, the electroneurographic (ENG) signals recorded from extracellular electrodes (e.g. cuff sensors and intrafascicular electrodes) measure the multi-unit activity from a population of axons in their immediate vicinity. The multi-unit activity conveys a mixture of time-overlapping single-unit activity (i.e., trains of action potentials) from different axons that corresponds to various motor commands. This information is coded in the action potential frequency and the nerve's cross-sectional firing distribution and is largely lost in multi-unit activity unless it is possible to recover single-unit spike trains from the interleaved multi-unit activity.

The central theme of the proposed neural decoding is to reconstruct the cross-sectional "spikiness" pattern of peripheral fascicules (corresponding to selective innervation of target muscles or motor units) based on a limited number of channels and noisy ENG signals available. For the cuff electrodes, we will develop algorithms to separate the active neural signals from the myoelectric interference, and for the intrafascicular electrodes we will to decode the neural spike or field potential activity recorded

Neuromorphic VLSI implementation

Reconfigurable neuromorphic chip that performs bio-realistic spiking neuronal and synaptic computations in 0.35 µm CMOS; and measured neuronal firing patterns from the IC.
The last step in the neural decoding project is the neuromorphic VLSI implementation of the decoding algorithms. Taking inspiration from the human brain which accomplishes similar tasks at amazing power efficiencies exceeding 1000 GMAC/J, we have developed reconfigurable neuro-mimetic integrated circuits that feature spiking neurons and synapses with bio-realistic properties. Recognized by the prestigious TR35@Singapore award by MIT's Technology Review, our neuromorphic approach can implement spiking neural networks at power efficiencies that are more than 1000x better than low-power digital processors.

Figures and text from these pages may not be reproduced or used for any purpose without permission from SINAPSE.