Fast and energy-efficient neuromorphic deep learning with first-spike\n times

Abstract

For a biological agent operating under environmental pressure, energy\nconsumption and reaction times are of critical importance. Similarly,\nengineered systems are optimized for short time-to-solution and low\nenergy-to-solution characteristics. At the level of neuronal implementation,\nthis implies achieving the desired results with as few and as early spikes as\npossible. With time-to-first-spike coding both of these goals are inherently\nemerging features of learning. Here, we describe a rigorous derivation of a\nlearning rule for such first-spike times in networks of leaky\nintegrate-and-fire neurons, relying solely on input and output spike times, and\nshow how this mechanism can implement error backpropagation in hierarchical\nspiking networks. Furthermore, we emulate our framework on the BrainScaleS-2\nneuromorphic system and demonstrate its capability of harnessing the system's\nspeed and energy characteristics. Finally, we examine how our approach\ngeneralizes to other neuromorphic platforms by studying how its performance is\naffected by typical distortive effects induced by neuromorphic substrates.\n