The dynamic interplay between in-context and in-weight learning in humans and neural networks

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Neuroscience

Background:

• Human learning exhibits a duality: rule-based and incremental. Psychological theories propose distinct learning systems for these behaviors.
• Neural networks traditionally model incremental learning via weight updates, struggling to explain rule-based learning.

Purpose of the Study:

• To investigate if neural networks capable of in-context learning (ICL) can replicate human learning duality.
• To reconcile dual-process theories of human learning with neural network capabilities.

Main Methods:

• Utilized metalearning neural networks and large language models capable of in-context learning (ICL).
• Evaluated network performance on rule-governed and unstructured tasks, comparing with human behavioral data.
• Differentiated between in-context learning (ICL) and in-weight learning (IWL) mechanisms.

Main Results:

• Networks with ICL demonstrated human-like compositional learning on rule-governed tasks.
• These networks also replicated human trial-and-error learning on unstructured tasks via IWL.
• Emergent ICL in neural networks provides distinct learning properties alongside IWL.

Conclusions:

• In-context learning (ICL) equips neural networks with flexible learning abilities, bridging the gap between rule-based and incremental learning.
• This reconciles dual-process theories with neural network models, offering new perspectives on cognitive flexibility.
• ICL and IWL can coexist in neural networks, mirroring human dual learning systems.