A biologically plausible sparse approximation solver on neuromorphic hardware

Abstract

We develop a novel design methodology to map the biologically plausible Locally Competitive Algorithm (LCA) to the brain-inspired TrueNorth chip to solve for the sparse approximation of a signal, offering the largest LCA dictionaries implemented on neuromorphic hardware to date with perfect precision. We observe low-power consumption in the operation of the LCA on the TrueNorth chip. We also explain methods to map other sparsity-based probabilistic inference problems onto the hardware using our design methodology. We describe the optimal way to achieve high-precision calculations by encoding and decoding signals within time windows. We discuss in detail functional processing units for use on the hardware that offer non-linear thresholds, increased vector-matrix multiplication precision, and the ability to accurately implement a recurrent network on the TrueNorth chip. Our design methodology offers the foundation for low-power embedded systems signal processing applications using the TrueNorth chip.