A scalable time-based integrate-and-fire neuromorphic core with brain-inspired leak and local lateral inhibition capabilities

Abstract

A fully scalable light-weight integrate-and-fire neuromorphic core with brain-inspired leak and local lateral inhibition features is implemented in 65nm. The core computes the neural net algorithm entirely in the time domain using standard digital circuits. A parallel two-layer architecture realized using the proposed core achieves a 91% handwritten digit recognition accuracy. The 0.24mm2 neuromorphic core including 64 digitally controlled oscillator (DCO) circuits consumes 320.4 W per DCO at a maximum throughput of 746M pixels/s.