Séminaire interne - Francesca Schönsberg

Title: 

Perceptual biases emerge in a recurrent neural network with ongoing Hebbian plasticity

Abstract:

Perceiving the magnitude of a stimulus is a complex brain function that arises from the interplay of working memory and experience. This interplay results in two well-known perceptual biases in memory tasks:
1. In a series of vibrational stimuli of varying strength, both humans and rodents tend to overestimate the strength of a stimulus after a series of weak stimuli (and vice versa) due to the repulsive bias of
representations. 2. The contraction bias instead shifts the representation of a stimulus held in working memory towards the average of stimuli observed in the past. While a series of experiments have yielded a
detailed phenomenological description of both biases, the neural mechanisms underlying these biases remain poorly understood.

In the presentation I will report a recent study in which  we show that the representations learnt by recurrent neural networks with ongoing Hebbian plasticity quantitatively reproduce (i) the contractive bias we find in experiments with human subjects, and (ii) the repulsive effect found in rodents by Hachen et al. (Nat. Comm. 2021). In our model, a fully-connected  network of rate-based units is driven by external inputs modeled after the experimental protocol, while its connectivity is continuously evolving due to Hebbian plasticity. We do not use gradient descent nor do we fine-tune the model to different experimental paradigms. We finally design a new behavioural paradigm where contraction and repulsive bias interact and find again that the model predicts salient features of the performance of our human participants.

Our results show that a single recurrent neural network with ongoing Hebbian plasticity reproduces two perceptive biases observed across three experimental paradigms. The striking match between experimental data and theoretical predictions supports the hypothesis that perceptual biases arise from simple Hebbian plasticity within a unique recurrent subregion of the brain, e.g. vM1 in rats, which acts as a plastic platform that filters perception based on context.