RESEARCH DIRECTION
Pure pattern recognition interpolates; intelligence must extrapolate. Deep learning's well-documented failure modes — data hunger, brittleness under distribution shift, shallow compositionality — are not engineering accidents but architectural signatures. We build hybrid systems in which learned representations propose and explicit symbolic machinery composes, checks, and proves.
The division of labor is principled: neural components handle perception, retrieval, and hypothesis generation — the tasks where statistical learning excels — while symbolic components handle binding, composition, and inference over abstract variables, where systematicity is non-negotiable. Neither half is decoration; each covers the other's documented failure modes.
A system that has learned 'red cube' and 'blue sphere' should handle 'blue cube' at zero marginal cost. We treat systematic recombination — operations over roles and variables rather than memorized configurations — as a designed-in property, measured by held-out compositions the system has never seen, not by in-distribution accuracy.
When reasoning chains are explicit, they can be checked. We favor architectures whose conclusions come with derivations — inference traces that can be audited step-by-step, formally verified where the domain permits, and rejected when a step fails. Trust in the answer reduces to trust in the checker, which is exactly where trust belongs.
WORKING PRINCIPLES
CONTINUE EXPLORING