State of the art machine learning relies on the backpropagation of errors to train artificial neural networks, as this learning algorithm has proven to be the most powerful in terms of accuracy on complex tasks. However, backpropagation is not adapted to hardware neural networks. Indeed it requires circuitry to explicitly compute error gradients over the network, as well as to update the synaptic weights. It is highly non-local, as computing weight updates in one layer might require knowledge of the state of layers at the over end of the network. Furthermore, the required weight update during learning are small and thus hard to achieve with imperfect physical devices.

On the contrary, learning in our brain is highly local and emerges from neural dynamics. Although the exact mechanism of learning in the brain is elusive, many brain-inspired learning algorithms have been proposed. On top of their bio-plausibility, they are better adapted to hardware artificial neural networks than backpropagation. However, they struggle to reach high accuracy for complex tasks.

An exciting topic is thus to develop novel learning algorithms, merging the accuracy of backpropagation and the hardware-friendlyness of the brain-inspired algorithms. We think that our physical neural networks are a natural platform to implement algorithms which achieve computing and learning through the dynamics of the components. Therefore we co-design dynamical neural network hardware and algorithms.

Key results and publications:

E. Martin, M. Ernoult, J. Laydevant, S. Li, D. Querlioz, T. Petrisor, and Julie Grollier, EqSpike: Spike-driven equilibrium propagation for neuromorphic implementations, iScience 24, 3 (2021)

J. Laydevant, M. Ernoult, D. Querlioz, J. Grollier, Training Dynamical Binary Neural Networks with Equilibrium Propagation, Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021)

M. Ernoult, J. Grollier, D. Querlioz, Y. Bengio, B. Scellier, Equilibrium Propagation with Continual Weight Updates, [arXiv] (2020)