An autoassociative neural network model of paired-associate learning

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Machine Learning

Background:

• Hebbian heteroassociative learning typically creates asymmetric associations (A to B, but not B to A).
• Recurrent neural networks offer a solution by enabling bidirectional associations (A to B and B to A).
• Differential weighting of forward and backward associations in recurrent networks can model recall asymmetries.

Purpose of the Study:

• To analyze a general recurrent neural network model of associative memory.
• To evaluate the model's ability to fit experimental data on human associative learning.
• To investigate the relationship between forward and backward storage strengths in associative recall.

Main Methods:

• Developed and analyzed a recurrent neural network model for associative memory.
• Compared model fits to human associative learning data under varying correlations of forward and backward storage strengths.
• Utilized computational modeling to interpret psychological data on associative recall.

Main Results:

• The recurrent network model provided a significantly better fit to human associative learning data when forward and backward storage strengths were highly correlated.
• This suggests that a single, unified representation (blending) is a more effective model than separate forward and backward associations.
• The findings highlight the importance of the correlation between association strengths in explaining recall performance.

Conclusions:

• Human associative learning is better conceptualized as a blending of two ideas into a single memory unit.
• Recurrent neural network models support this unified representation view over separate, independently modifiable associations.
• The correlation between forward and backward learning strengths is crucial for understanding associative recall dynamics.