NeuPI v1.0.1
NeuPI v1.0.1 Release Notes
With minor bug fixes
We are thrilled to release NeuPI v1.0.1. This version marks a major milestone, transitioning the library from a collection of core evaluators into a full-fledged framework for advanced neural probabilistic inference. This release introduces powerful new inference schemes, sophisticated discretization methods, and significant improvements to the library's core architecture, making it more modular, extensible, and robust.
✨ New Features
Advanced Inference Schemes
ITSELF_Engine: The highlight of this release is the introduction of the Inference Time Self-Supervised Learning Fine-tuning engine. This advanced inference module performs on-the-fly optimization for each test instance or batch, refining the neural model's parameters to significantly improve solution quality over a standard single-pass approach.
Sophisticated Discretization Methods
The discretize module has been expanded with powerful methods that go far beyond simple thresholding:
KNearestDiscretizer: Implements a high-performance beam search to find the k-nearest binary vectors to the network's continuous output. It uses a PGM evaluator as a scoring function to identify high-quality discrete solutions, backed by an optimized Cython helper for performance.HighUncertaintyDiscretizer: A new heuristic search method that focuses computational effort where it matters most. It identifies the k query variables with probabilities closest to 0.5 (the highest uncertainty) and performs an exhaustive search over this reduced space to find the best assignment.
Modular Preprocessing
preprocessModule: Introduced a dedicated module for feature engineering. This separates the creation of model inputs from the model architecture itself.
This release lays a robust foundation for future research and development. We look forward to expanding the library's capabilities in upcoming versions.