We introduce BiMU, a Bayesian update rule for binary neural networks that prevents posterior saturation on long non-stationary streams, sustaining epistemic uncertainty and enabling buffer-free active querying to cut label and backpropagation costs.
Current research
My thesis is about Bayesian metaplasticity. Each weight of a network keeps a measure of its own uncertainty, and that uncertainty decides how much the weight may still move: what to remember, and what to forget. With Djohan Bonnet we turned this into an update rule that learns and forgets without being told where one task ends (Nature Communications, 2025). I then carried it over to binary weights, where a Bayesian network otherwise saturates and stops learning, so that it can run on hardware that stores a single bit per synapse (ICML 2026). The thesis also exists as a four-and-a-half-minute animation, with its transcript.
The work is part of the Metaspin project, which builds energy-efficient computing systems inspired by the brain, funded by the European Innovation Council (EIC). Along the way I have helped the device physicists at C2N turn magneto-ionic devices into synapses.
Latest paper all publications
Latest writing all posts
Two Bayesian continual-learning rules, one term apart, and what that term does to a posterior on a stream that refuses to sit still.
Talks & appearances all talks
- Oct 2026PhD defence, Palaiseau
Bayesian metaplasticity. The video is a 4½-minute animated summary of the thesis. video
- Jul 2026ICML 2026, Seoul
Poster. Active continual learning with metaplastic binary Bayesian neural networks slidesposter
- Sep 2025Neuromorphic Symposium, Paris
Talk, 15 min. Bayesian continual learning and forgetting in neural networks slides