Recurrent networks endowed with structural priors explain suboptimal animal behavior

Area of Science:

• Neuroscience
• Computational Neuroscience
• Animal Behavior

Background:

• Animal strategies in novel tasks blend individual experience with evolved structural priors.
• Rats in two-alternative forced choice (2AFC) tasks learn from correct trials but err after incorrect ones.

Purpose of the Study:

• Investigate the neural mechanisms behind outcome-dependent learning in rats.
• Understand how evolved priors influence decision-making strategies in laboratory tasks.

Main Methods:

• Compared rat behavior in 2AFC tasks with recurrent neural networks (RNNs) trained on the same task.
• Pre-trained RNNs on an ecological task to mimic evolved priors and observed their behavior in 2AFC tasks.
• Utilized population analyses to examine network dynamics and evidence accumulation.

Main Results:

• RNNs trained solely on 2AFC tasks outperformed rats by using across-trial information after both correct and error trials.
• RNNs pre-trained on an ecological task developed error-specific gating, mirroring rat behavior.
• Error trials in pre-trained networks induced dynamic changes, decoupling stimulus categorization from accumulated evidence.

Conclusions:

• Rats' suboptimal performance in 2AFC tasks stems from a structural prior adapted to natural environments.
• This prior imposes an error-driven gating mechanism, isolating decision dynamics and limiting task performance.
• Computational models can reveal how evolved priors shape behavior in simplified laboratory settings.