Adaptive Learning through Temporal Dynamics of State Representation.

Area of Science:

• Cognitive Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Human learning rates adapt to environmental statistics.
• Existing theories lack comprehensive explanations for adaptive learning behaviors and their neural underpinnings.
• Understanding how individuals adjust learning sensitivity in dynamic environments is crucial.

Purpose of the Study:

• To develop a neural network model explaining adaptive learning rate adjustments.
• To provide a mechanistic interpretation of neural correlates of learning, including P300 and pupil dilation.
• To offer a unified framework for understanding learning across diverse statistical contexts.

Main Methods:

• A feedforward neural network model was developed.
• The model learns to map internal context representations to behavioral outputs via supervised learning.
• Internal state transitions were manipulated to modulate learning rates based on feedback mismatch duration.

Main Results:

• The model demonstrates that sustained state transitions after changepoints enhance learning, mimicking neural reset phenomena.
• Short-lived state transitions following outlier events limit their impact on future behavior.
• Model dynamics provide a mechanistic account for bidirectional learning signals and their relation to surprise.

Conclusions:

• Dynamic latent state representations enable normative inference in learning.
• The model offers a coherent framework for understanding neural signatures of adaptive learning.
• This approach bridges normative learning theories with latent state representation, providing testable predictions.