Research & Papers

PyVaCoAl architecture fulfills Marcus's 25-year-old algebraic mind vision

XOR-and-shift hyperdimensional computing achieves reversible variable binding and compositionality

Deep Dive

Twenty-five years after Gary Marcus outlined three essential components for any adequate cognitive architecture—operations over variables, recursively structured representations, and a distinction between individuals and kinds—researchers Hiroyuki Chuma, Kanji Otsuk, and Yoichi Sato have delivered a concrete implementation. Their PyVaCoAl/VaCoAl system is a hyperdimensional computing architecture built on a single algebraic primitive: XOR-and-shift over Galois field GF(2), executed via primitive-polynomial linear-feedback shift registers. This substrate achieves reversible variable binding (Bind(R,F) = R XOR shift(F)) and non-commutative compositional bundling that can distinguish 'the dog bites the man' from 'the man bites the dog.' The architecture naturally separates address spaces for individuals versus kinds under the same algebra, directly addressing Marcus's third pillar.

Crucially, the paper maps this architecture to biological circuitry, arguing that the dentate gyrus-CA3 circuit is a functional homologue, with developmentally specified mossy-fiber targeting providing the innate microcircuitry Marcus anticipated. Unlike earlier attempts using tensor products, circular convolution, or temporal synchrony, PyVaCoAl/VaCoAl offers a far tighter match to Marcus's specifications. It also naturally extends to Judea Pearl's rung-3 counterfactual reasoning—a capability the original treelet proposal did not target. This work represents a significant step toward a neural substrate that combines symbolic reasoning with biologically plausible implementation.

Key Points
  • PyVaCoAl uses XOR-and-shift over GF(2) as its single algebraic primitive, implemented via primitive-polynomial linear-feedback shift registers.
  • It achieves reversible variable binding (Bind(R,F)=R XOR shift(F)) and non-commutative bundling, distinguishing compositional structures like 'dog bites man' vs 'man bites dog'.
  • The architecture maps to Marcus's three pillars and extends to Pearl's rung-3 counterfactual reasoning, surpassing earlier approaches like tensor products or circular convolution.

Why It Matters

This architecture offers a biologically plausible neural substrate for symbolic reasoning, bridging neuroscience and AI.