François Danneville

François Danneville, IRCICA, Université de Lille

Scaling progress of an ultra-low power neuromorphic technology for spiking neural network

The aim of this talk is to address the progress made to scale an ultra-low power (ULP) neuromorphic 65 nm CMOS technology for spiking neural network (SNN) design, through the presentation of various designed circuits operating under few hundreds mV supply voltage. In a first part, the artificial neuron with its different variants and features will be presented, followed by a complete toolbox dedicated to SNN design, gathering excitatory/inhibitory plastic synapses as well as an ULP optical sensor. In a second part, the usefulness of this toolbox will be highlighted in the framework of various contexts (among them, circuits allowing the emulation of burst firing or the detection of direction/speed). The last part will focus on results achieved on a medium scaled SNN (few tens of neurons) allowing recognition of oriented edges (3*3 input pixels) through a supervised learning.