Comparing continual task learning in minds and machines.

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Machine Learning

Background:

• Humans demonstrate remarkable continual learning abilities over a lifespan, a feat currently challenging for artificial intelligence (AI).
• Existing machine learning systems often exhibit catastrophic forgetting when learning new tasks sequentially.
• Understanding human cognitive mechanisms is key to developing more robust AI.

Purpose of the Study:

• Investigate cognitive mechanisms enabling human continual learning.
• Apply behavioral findings to design improved neural network architectures.
• Address catastrophic forgetting in deep learning models.

Main Methods:

• Human participants learned to categorize tree images using two orthogonal task rules via trial and error.
• Compared blocked training (prolonged focus on single rules) with interleaved training.
• Analyzed human error patterns to infer representational strategies.
• Trained standard deep neural networks (DNNs) and augmented DNNs with unsupervised generative models.

Main Results:

• Blocked training improved human performance on interleaved rule tests, suggesting the formation of factorized representations.
• Standard DNNs exhibited catastrophic forgetting under blocked training due to representational interference.
• Augmenting DNNs with unsupervised generative models to learn stimulus embeddings reduced catastrophic forgetting.

Conclusions:

• Factorized representations, facilitated by blocked training, are crucial for human continual learning.
• Unsupervised pre-learning of world embeddings can mitigate catastrophic forgetting in AI.
• Developing artificial agents that first model the world is a promising direction for AI continual learning.