A Computational Model of Context-Dependent Encodings During Category Learning

Area of Science:

• Cognitive Psychology
• Computational Neuroscience
• Machine Learning

Background:

• Exemplar models of category learning offer flexibility but struggle with sequential learning effects.
• Existing models do not adequately adapt to local changes in study sequences, such as interleaved vs. blocked trials.

Purpose of the Study:

• Introduce the Sequential Attention Theory Model (SAT-M) to address limitations in current category learning models.
• Investigate how local context, specifically temporal item relationships, influences category acquisition.
• Compare SAT-M's predictive power against established models like ALCOVE and SUSTAIN.

Main Methods:

• Developed SAT-M, incorporating both global (category) and local (temporal) context into item encoding.
• Fitted SAT-M to experimental data from three studies comparing interleaved and blocked training sequences.
• Analyzed model parameters and compared predictions with behavioral data, including learners' looking times.

Main Results:

• SAT-M accurately captured the impact of local context on category learning across different training sequences.
• The model successfully predicted performance differences between interleaved and blocked training conditions.
• Established models (ALCOVE, SUSTAIN) and a simplified SAT-M showed poorer fits to the experimental data.

Conclusions:

• SAT-M provides a more comprehensive account of category learning by integrating sequential effects.
• The model's success highlights the importance of local context in understanding how training order impacts learning.
• Empirical validation through looking times supports SAT-M's predictions about encoding changes due to training sequences.