Dr. Kellian Cottart

Bayesian metaplasticity

Uncertainty as a computational principle for resource-constrained learning. PhD thesis, Université Paris‑Saclay (C2N), defended 5 October 2026.

A network trained on a stream that keeps changing can fail in two ways: it overwrites what it learned before, or it holds on to it so firmly that nothing new gets in. In my thesis, each synapse carries a measure of its own uncertainty, and that uncertainty decides how far the synapse may still move.

Transcript · the captions on screen; click a line to play from there

Where each part comes from

The animation follows the order of the defence: the update rule for Gaussian synapses (MESU), the active learning that the same uncertainty makes possible, the rule for binary synapses (BiMU), and a magneto-ionic synapse whose updates a magnetic field can shape. The papers and essays below have the derivations and the full results that the animation leaves out.

  1. The idea

    Synaptic plasticity as Bayesian inference

    Aitchison et al., Nature Neuroscience, 2021. The neuroscience model in which each synapse tracks its own uncertainty.

  2. MESU

    Bayesian Continual Learning and Forgetting in Neural Networks

    Nature Communications, 2025. Gaussian synapses that forget slowly toward the prior. The essay When Bayesian neural networks remember too much follows the variance collapse that the video shows in one sentence.

  3. Active learning, BiMU

    Active Continual Learning with Metaplastic Binary Bayesian Neural Networks

    ICML, 2026. Binary synapses with a memory window, and the variation ratio that decides which images are worth a label. Also as a post and an ICML poster.

  4. Hardware

    Dynamic Control of Weight-Update Linearity in Magneto-Ionic Synapses

    Nano Letters, 2025. A magnetic field applied to the device tunes how linear its depression is, meaning how much the size of each step depends on the weight it starts from; the video shows this as the field setting the size of each step. Followed by Differential Magneto-Ionic Artificial Synapse (Appl. Phys. Lett., 2026).