SKM 2021 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

SYNC: Symposium Advanced neuromorphic computing hardware: Towards efficient machine learning

SYNC 1: Symposium: Advanced neuromorphic computing hardware: Towards efficient machine learning

SYNC 1.1: Invited Talk

Wednesday, September 29, 2021, 10:00–10:30, Audimax 1

Equilibrium Propagation: a Road for Physics-Based Learning — •Damien Querlioz — Université Paris-Saclay, CNRS, C2N, Palaiseau, France.

Neuromorphic computing takes inspiration from the brain to create highly energy-efficient hardware for information processing, capable of sophisticated tasks. The resulting systems are most often preprogrammed: training neuromorphic systems on-chip to perform new tasks remains a formidable challenge. The flagship algorithm for training neural networks, backpropagation, is indeed not hardware-friendly. It requires a mathematical procedure to compute gradients, external memories to store them, and an external dedicated circuit to change the neural network parameters according to these gradients. The brain, by contrast, does not learn this way. It learns intrinsically, and its synapses evolve directly through the spikes applied by the neurons they connect, using their biophysics. This technique is very advantageous in terms of energy efficiency and device density. In this talk, I will introduce our approach towards reproducing this brain strategy of intrinsic learning exploiting device physics. I will show through simulations how we take advantage of the physical roots of an algorithm called Equilibrium Propagation (1) to design dynamical circuits that learn intrinsically with high accuracy (2-4).

1. B. Scellier, Y. Bengio, Front. Comput. Neurosci. 11 (2017). 2. M. Ernoult, J. Grollier, D. Querlioz, Y. Bengio, B. Scellier, Proc. NeurIPS, pp. 7081 (2019). 3. A. Laborieux et al., Front. Neurosci. 15 (2021). 4. E. Martin et al., iScience. 24 (2021).