Rapid learning of temporal dependencies at multiple timescales

Area of Science:

• Cognitive Science
• Computational Neuroscience

Background:

• The human brain processes complex environmental information across various timescales.
• Understanding how the brain learns and represents simultaneous temporal information streams is crucial for cognitive modeling.

Approach:

• A statistical learning paradigm was employed with human participants (N=96) and a gated recurrent neural network.
• Participants learned sequential target location orders at short and long timescales within a game.
• A recurrent neural network was trained to predict target locations, mirroring human learning processes.

Key Points:

• Human participants demonstrated context-dependent learning of order information at both short and long timescales.
• Learning was more pronounced for short-timescale contingencies compared to long-timescale ones.
• The computational model exhibited multilevel learning and replicated human sensitivity to the games' similarity structure.

Conclusions:

• Both humans and artificial neural networks can rapidly acquire and represent order information concurrently across different timescales.
• The findings suggest a shared mechanism for processing hierarchical temporal information in biological and artificial systems.
• This research provides insights into the neural basis of statistical learning and temporal information processing.