Talk abstracts > Florent Meyniel (Neurospin, Saclay)

Go back to Homepage.

Florent Meyniel - Neurospin, Paris-Saclay

Thursday, June 1^st

Talk Session 1: How artificial neural networks make decisions

10h - 10h30

Adaptive learning in brains and machines

Natural environments are both stochastic and dynamic. These two aspects should promote opposite components in learning, namely stability, to maintain estimates in the face of stochastic noise, and flexibility, to quickly adapt estimates when the environment changes. Formally, this trade-off is controlled by the learning rate, which is dynamically adjusted in adaptive learning algorithms. Human learning has been shown to be adaptive, especially when learning magnitudes (e.g., reward size). I will show that human learning of probabilities (e.g. of stimulus occurrence) is also adaptive. A normative analysis shows that the adaptation of learning rates should be different when learning magnitudes and probabilities: the modulation of learning rates by uncertainty should be more pronounced for probability learning. In practice, this uncertainty can be reported by subjects and it modulates their learning rates. In the same task, artificial neural networks (RNNs) discover this uncertainty weighting principle when trained with self-supervision. RNNs automatically develop a representation of uncertainty and use it to regulate their learning rate when their connections are equipped with a gating mechanism. Together, these results demonstrate the extent of uncertainty-based adaptive learning in humans and provide insight into its possible neural implementation.