公开(公告)号：GB2548194A

公开(公告)日：2017-09-13

申请号：GB201700228

申请日：2017-01-06

Applicant: IBM

Inventor： AMON AMIR ,  ANTONIO JOSE JIMENO YEPES ,  JIANBIN TANG

IPC: G06N3/04 ,  H03M7/02

Abstract: Weighted population code for efficient data representation in neuromorphic systems. Input value is represented by spikes on axons, each of which has an associated weight. Weighted spikes are integrated/added, which may be over over multiple ticks, to give the input value. Requires less hardware for greater dynamic range compared to prior art. Can be converted to stochastic code by dividing value by size of dynamic range. Also disclosed is stochastic code corelet 206 on neuromorphic chip 205. Pre-processor 202 encodes input data to weighted population code for transfer onto neuromorphic chip, and corelet translates input to stochastic code for use by classifiers also on chip. This reduces the input bandwidth of the chip. Also disclosed is delaying sending of spikes to a neuron during a preload period. The spikes then affect the neuron during a subsequent setup period, before the neuron produces output. This means that input can be sent while the neuron is still producing the previous output, which reduces setup time and shortens processing cycle.