Identifying Transfer Learning in the Reshaping of Inductive Biases.

Area of Science:

• Cognitive Science
• Neuroscience
• Artificial Intelligence

Background:

• Transfer learning, crucial for human intelligence, involves reusing knowledge in new situations.
• Computational mechanisms of human transfer learning, particularly inductive bias updating, remain under-investigated.
• Efficient inductive biases generalize prior experiences, shaping learning through meta-level constraints.

Purpose of the Study:

• To investigate how humans update inductive biases for effective transfer learning.
• To explore the role of internal models in adapting to changing task structures.
• To determine if subjective internal models predict cross-task transfer performance.

Main Methods:

• Participants trained on a visual sequence task (Alternating Serial Response Times - ASRT).
• Training involved exposure to a specific sequence over multiple days.
• Transfer phase introduced a changed sequence while maintaining the underlying task structure.

Main Results:

• Participants updated their inductive biases beyond sequence acquisition.
• Prior exposure accelerated learning of new sequences, especially in individuals abandoning initial biases.
• Learning enhancement correlated with developing new internal models and alternating between them.

Conclusions:

• Humans dynamically update inductive biases, enabling efficient transfer learning.
• Subjective internal models are key predictors of successful transfer across tasks.
• Imperfect prior learning aids new learning by leveraging partial knowledge of regularities.