Encoding sequential information in semantic space models: comparing holographic reduced representation and random...

Area of Science:

• Cognitive Science
• Neuroscience
• Computational Linguistics

Background:

• Binding operators are crucial for encoding sequential information in semantic memory.
• Circular convolution and random permutation are proposed neurally plausible binding operators.
• Comparing their efficacy in memory encoding is essential for understanding semantic representation.

Purpose of the Study:

• To compare circular convolution and random permutation as binding operators for semantic memory.
• To evaluate their performance in encoding paired associates and sequential information.
• To assess their neurological plausibility and utility in vector space models.

Main Methods:

• Controlled experimental comparisons of circular convolution and random permutation.
• Evaluation of paired associate learning and sequential information encoding.
• Analysis of performance across different corpus sizes and permutation types (true vs. noisy).

Main Results:

• Random permutations significantly outperformed circular convolution in storing paired associates.
• Performance was comparable on small corpora but favored random permutations for large corpora due to scalability.
• Noisy permutations demonstrated performance close to true permutations, enhancing neurological plausibility.

Conclusions:

• Random permutations are a more effective and scalable binding operator than circular convolution for semantic memory.
• The findings support the neurological plausibility of random permutations, including noisy variants.
• Random permutations show significant utility in developing advanced vector space models of semantics.