Novelty based on representational similarity

Original Paper: Representational similarity modulates neural and behavioral signatures of novelty

Introduction

Definition of novelty

Novelty is an information-gathering mechanism that can serve as a metric for the amount of newly acquired information. Novelty is closely related to surprise, which corresponds to the discrepancy between a prediction and the information actually obtained. These mechanisms are critical for an agent to adapt to non-stationary and uncertain environments.

Previous studies

Previous studies of novelty have mostly focused on discrete, count-based measures. However, because stimulus spaces are often continuous, novelty—or, equivalently, newly acquired information—can generalize across similar stimuli. Such generalization has not been adequately accounted for. This paper therefore proposes a theoretical framework for generalizing novelty to continuous stimulus spaces.

Results

A new model for novel stimuli

Different visual stimuli can lie within a continuous stimulus space, such as orientation in a visual field. Count-based novelty cannot sufficiently distinguish between two sequences of stimuli that contain different amount of information. The authors construct a formula for a novelty signal that better captures the intuitive notion of novelty.

Novelty explains visual cortex activity

Using the proposed novelty formula, the authors successfully explain visual cortex activity in response to different visual stimuli. Gabor stimuli (a series of oriented stripes) is presented to a mouse, and neural activity in the visual (V1) cortex is analyzed. The new model better predicts the activity compared to the count based model.

Novelty guides behavioral exploration

The authors attempt to show that the novelty signal also guides exploratory behavior. They analyze mouse behavior in a maze exploration task. The observed behavior is successfully explained by a novelty-seeking RL algorithm. Although the specific choice of novelty signal is somewhat arbitrary, the data support the broader concept of novelty explaining exploration.

Novelty framework

The authors propose a novelty framework that addresses the discrimination–generalization trade-off between the recency effects of stimuli and the propagation of novelty across stimulus spaces. They also predict the results of an experiment designed to validate the specific form of their novelty paradigm.

Discussion

The paper investigates the gap between generalization and discrimination from the perspective of novelty. Several limitations remain, including the arbitrary choice of representation and the limited generalizability of some of the paper's claims. Nevertheless, the proposed concepts and approaches could be applied to other research topics and may provide a useful new perspective on representation.