The abstract is read at the publisher; the summary is JClub's.
In closed-loop sensory environments, simple networks with local Hebbian-like plasticity learn reward-seeking behaviors without any reward objective, with prospective configuration matching multiple phenomena.
- Why it matters: Reward-seeking is often treated as reward optimization, but it is unclear how adaptive behavior can emerge from local learning rules that lack explicit reward signals.
- What they did: Six local learning rules were tested in closed-loop setups where rewards were provided as ordinary sensory input, and only prospective configuration robustly reproduced reward-magnitude-dependent learning rates, probability matching, context-retained memory, and exploration–exploitation switching with heavy-tailed statistics.
- The result: Network actions that generate input become self-sustaining attractor states, whose activity–weight geometry adapts to the environment, enabling a single framework to predict behavior and neural population geometry changes.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.