Meta-learned models of cognition

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Artificial Intelligence

Background:

• Computational models are crucial for psychology and neuroscience.
• Traditional models like cognitive architectures and Bayesian models are hand-designed.
• Meta-learning offers a data-driven alternative for building cognitive models.

Purpose of the Study:

• To synthesize existing research on meta-learned models of cognition.
• To establish a coherent research program for meta-learning in cognitive science.
• To highlight the advantages of meta-learning over traditional modeling approaches.

Main Methods:

• Review and synthesis of existing literature on meta-learning in cognitive science.
• Demonstration of meta-learning's ability to construct Bayes-optimal learning algorithms.
• Analysis of meta-learning's connection to the rational analysis of cognition.

Main Results:

• Meta-learning models acquire inductive biases from environmental interactions.
• Meta-learning can be used to create Bayes-optimal learning algorithms.
• Meta-learning offers advantages over hand-designed computational models.

Conclusions:

• Meta-learning presents a promising framework for developing computational models of cognition.
• This approach strengthens the link between computational models and the rational analysis of cognition.
• A unified research program for meta-learned cognitive models is proposed.