Compositional pretraining improves computational efficiency and matches animal behavior on complex tasks.

Area of Science:

• Computational Neuroscience
• Machine Learning in Biology

Background:

• Recurrent neural networks (RNNs) are widely applied in neuroscience to model neural dynamics and behavior.
• Traditional RNN training methods struggle with complex cognitive tasks and capturing nuanced animal behaviors.

Purpose of the Study:

• To develop a principled approach for incorporating compositional tasks into RNN training.
• To enhance RNNs' ability to model complex cognitive behaviors, specifically in a temporal wagering task studied in rats.

Main Methods:

• Designed a pretraining curriculum using simpler cognitive tasks reflecting sub-computations relevant to the target task.
• Trained RNNs on this curriculum before tackling the complex temporal wagering task.

Main Results:

• Pretraining significantly improved RNN learning efficacy.
• RNNs trained with this method adopted strategies similar to rats, including long-timescale inference of latent states.
• Conventional pretraining methods failed to capture these crucial aspects.

Conclusions:

• The proposed compositional pretraining approach endows RNNs with relevant inductive biases for modeling complex behaviors.
• This method facilitates the development of slow dynamical systems features essential for inference and decision-making in RNNs.