Modelling cognitive flexibility with deep neural networks.

Area of Science:

• Artificial Intelligence
• Cognitive Science
• Neuroscience

Background:

• Deep reinforcement learning (DRL) models achieve human-level performance on complex tasks.
• Current DRL models lack adaptability, converging to fixed solutions unlike humans.

Purpose of the Study:

• To highlight novel methods enabling neural networks to model human cognitive flexibility.
• To explore how neural networks can adapt to new challenges and environments.

Main Methods:

• Training neural networks with complementary 'habit' and 'goal'-based policies.
• Meta-learning flexibility during pre-training on diverse data for in-context adaptation.
• Meta-training deep networks to adjust behavior based on environmental control levels.

Main Results:

• Neural networks can be trained to exhibit adaptive behaviors similar to human cognitive flexibility.
• Meta-learning approaches allow networks to rapidly adapt to novel situations.
• Generative models trained with reinforcement learning from human feedback offer new insights into cognitive flexibility.

Conclusions:

• Emerging methods allow neural networks to overcome fixed-solution limitations, mimicking human adaptability.
• These advancements pave the way for more flexible and human-like artificial intelligence.
• Further research with large generative models can deepen our understanding of cognitive flexibility.