Efficient Variational Inference in miniKanren with Weighted Model Countin...
summary
Extension of miniKanren implements probabilistic primitives and uses weighted model counting to compute likelihoods and EM-style variational updates, with a concrete Smalltalk implementation on GitHub. A naive search explodes (14-point k‑means becomes intractable) but inserting aggregate streams reduces complexity from exponential to linear and yields runtimes within an order of magnitude of hand‑written code (manual is about 4x faster in experiments). A third change, tabling tweaks, restores Viterbi/forward‑backward/Baum‑Welch style DP and makes HMMs and PCFGs efficient.
AI-written summary · about