Type-Preserving, Dependence-Aware Guide Generation for Sound, Effective Amortize...
summary
An automatic guide synthesizer generates neural-guide programs by a type-preserving, dependence-aware translation that guarantees mutual absolute continuity with the model even under recursion and stochastic branching. It extracts conditional independencies via program dependence graphs, emits guides that sample in alternate orders to reduce unnecessary coupling, and proves type-based AC soundness. On benchmarks it converges faster than mean-field and LSTM baselines, producing significantly higher effective sample sizes on the upside‑down‑tree and improved recognition rates for the context‑free recursion task.
AI-written summary · about
chapters (4)
- 0:00Start
- 5:27Faithfulness matters: IS with unfaithful guides is ineffective Faithfulness means guides do not contain conditional independences not found in models
- 10:42Absolute Continuity matters Real challenge: branching & recursion
- 18:54Combo of faithfulness and parsimony is a win
related talks
19mProgrammable MCMC with Soundly Composed Guide Programs
28mAutomated transpilation of imperative to functional code using neural-guided program synthe
26mTowards controllable language generation
24mReward-Guided Synthesis of Intelligent Agents with Control Structures
01h 14mLatest trends in AI
38mProgram Synthesis: A Dream Realized?
32mSecurity Implications of Artificial Intelligence Synthesizers
46mGenerative AI: Say Goodbye to Deterministic Testing?