Learning to perform role-filler binding with schematic knowledge.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Artificial Intelligence

Background:

• Schema theory posits mental frameworks (schemata) organize world knowledge.
• Generalizing knowledge requires role-filler binding: associating specific entities with abstract roles.
• Previous models required explicit labels or seen fillers for role-filler binding.

Purpose of the Study:

• To investigate if networks with external memory can perform role-filler binding for arbitrary fillers without explicit labels.
• To determine if these networks can generalize to role-filler pairs violating training correlations.
• To interpret learned representations using neural decoding techniques.

Main Methods:

• Developing a model architecture incorporating external memory.
• Training the network on structured data to learn relationships.
• Testing the model's ability to recall arbitrary fillers for specified roles.
• Evaluating performance on novel and correlation-violating role-filler pairs.
• Applying neural decoding methods for interpretability.

Main Results:

• Networks with external memory successfully learned role-filler binding for arbitrary fillers.
• The model generalized to novel role-filler pairings, including those violating training correlations.
• Performance on novel pairs did not significantly degrade knowledge of training correlations.
• Neural decoding analyses provided insights into the learned representations.

Conclusions:

• External memory enables models to perform flexible role-filler binding, crucial for generalizing knowledge.
• This approach allows models to handle novel situations and abstract concepts beyond direct training.
• The findings contribute to understanding how artificial systems can acquire and generalize knowledge akin to human cognition.