Multitask learning over shared subspaces

Area of Science:

• Cognitive Science
• Machine Learning
• Neuroscience

Background:

• Understanding how humans learn multiple tasks simultaneously is crucial for cognitive science.
• Machine learning models offer computational frameworks to investigate learning mechanisms.
• The concept of 'shared subspaces' in learning is an area of active research.

Purpose of the Study:

• To investigate if learning is enhanced when tasks share a common subspace.
• To compare human multitask learning performance with computational models.
• To explore the role of representational capacity and Bayesian learning in transfer effects.

Main Methods:

• Defined pairs of learning tasks based on shared or non-shared subspaces using machine learning.
• Human subjects learned these tasks via a feedback-based approach.
• Compared human performance to a sequential Bayesian learning Neural Network model.

Main Results:

• Human learning performance was significantly boosted when tasks shared a common subspace.
• Positive correlations in task performance were observed for shared subspaces.
• Human performance aligned with a minimal capacity Bayesian Neural Network model, not with higher capacity or non-Bayesian networks.

Conclusions:

• The concept of shared subspaces provides a valuable framework for studying human multitask and transfer learning.
• Bayesian learning with minimal capacity appears key to observing transfer effects similar to humans.
• Further research can utilize this framework to experimentally probe learning transfer.