Rapid Learning of Temporal Dependencies at Multiple Timescales

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Machine Learning

Background:

• The environment presents temporal information across multiple, simultaneous timescales.
• Understanding how humans and artificial systems learn and represent these dynamic information streams is crucial.

Purpose of the Study:

• To investigate human statistical learning of simultaneous short and long timescale contingencies.
• To model these learning processes using a recurrent neural network.

Main Methods:

• A statistical learning game with human participants (n=96) involving target location prediction.
• Manipulation of short-timescale (sequential pairs) and long-timescale (first vs. second half) order properties.
• Modeling using a gated recurrent network trained for immediate predictions.

Main Results:

• Human participants demonstrated context-dependent sensitivity to order information at both timescales, with stronger learning for shorter timescales.
• The recurrent network model exhibited multilevel learning and mirrored human sensitivity to game similarity structures.
• The model successfully grouped games by rule structure and differentiated based on low-level order information.

Conclusions:

• Humans and computational models can rapidly and concurrently acquire order information across different timescales.
• This research highlights the capacity for multilevel temporal information processing in both biological and artificial systems.