Holographic Reduced Representation

Holographic reduced representation (HRR) is a method for encoding complex symbolic structures as fixed-size vectors using circular convolution and superposition, developed by Tony Plate as an extension of tensor-product encoding. Unlike tensor products, which require dimensionality that grows with the number of bound roles and fillers, HRRs compress binding information into the same vector space as the constituents themselves, using a binding operation (circular convolution) that is approximately invertible. A proposition like "RUN(JOHN)" is encoded by convolving the vector for JOHN with a role vector for AGENT, then adding the result to a memory trace that can hold multiple such propositions simultaneously. Decoding uses correlation — the approximate inverse of convolution — to query the trace for constituents or roles.

The technique enables neural networks to perform analogical reasoning, role-filler binding, and structured memory retrieval without explicit symbolic data structures. Its central tradeoff is noise: as more propositions are superposed in a single trace, decoding accuracy degrades, imposing a soft capacity limit that mirrors human working memory constraints.