Weighted Model Integration
summary
Weighted Model Integration (WMI) extends weighted model counting to SMT with linear arithmetic and polynomial weights and comes with concrete solvers: predicate-abstraction, XADD compilation, search-based conditioning and piecewise-polynomial message passing. Complexity results show the largest provably tractable class is treewidth 1 with logarithmic diameter while raising treewidth to 2 or making a chain of length n yields #P-hardness. A practical toolchain implements a collapsed inference scheme that reduces Bayesian model averaging for ReLU fully connected layers to WMI, runs WMI per sample, and on ResNet benchmarks (CIBER) improves NLL, accuracy and ECE relative to sampling baselines.
AI-written summary · about
chapters (7)
- 0:00Introduction to WMI
- 1:47Defining WMI and motivation
- 6:49WMI for probabilistic inference
- 9:30Algorithms and complexity
- 21:16Application: Bayesian learning
- 29:44Collapse inference scheme
- 35:25Q&A and discussion
related talks
48mProbabilistic Reasoning by First-Order Model Counting
11mProbabilistic Programming Meets Automata Theory: Exact Inference using Weighted Automata
59mVerification and Machine Learning
22mMachine Learning with Wasm (wasi-nn)
11mWhen Less Is More: A Technical Overview of LLMs and the Strength of Smaller Models
40mSparse models are fast models: Improving DNN inference performance by over 10X
01h 03mSparsifying Sums of Norms
14mSYMI: Efficient Mixture-of-Experts Training via Model and Optimizer State Decoupling