Bayesian neural networks allow artificial intelligence systems to account for uncertainty in their predictions, but training them requires substantial computational resources. Resistive memories (ReRAM), which can both store data and perform computations directly within electronic circuits, offer a promising solution to reduce this cost.

This PhD project aims to develop a new Bayesian learning method on ReRAM hardware, based on Langevin sampling. This approach efficiently explores neural network parameters by combining random perturbations with gradient estimates. The objective is to exploit the intrinsic physical variability of ReRAM devices to estimate these gradients without relying on conventional backpropagation. The possibility of reusing memory programming attempts will also be investigated to reduce energy consumption and device wear.

The work will combine algorithm development, simulations, and experiments on existing ReRAM circuits at CEA-Leti. Expected outcomes include a demonstration of Bayesian learning on real hardware and an assessment of its performance, accuracy, and energy efficiency compared with existing approaches.