Survey
Informed ML Survey
This note tracks the core taxonomy of informed machine learning through knowledge source, representation, and integration location, with emphasis on the distinctions that remain useful for later comparison.
Reading Notes
Published notes with equations, figures, and implementation hooks preserved
These pages are generated from the Markdown notes in the repository. They document my ongoing reading and current understanding, with emphasis on mechanism, notation, and code-facing questions rather than final claims.
Logic Rules
Logic-Net
A rule-regularized teacher distribution is constructed first and then distilled back into student parameters, making the note useful for understanding posterior shaping, distillation, and optimization-level constraint injection.
Constraint Loss
Semantic Loss
This note centers on the 2018 Semantic Loss formulation where a propositional constraint is turned into a trainable loss through the probability mass of satisfying Boolean assignments.
Declarative Constraints
DL2
This note focuses on how DL2 turns logical constraints into a unified declarative interface that can shape both training and querying rather than only a single loss term.
Granular Computing
Granular Computing for ML
This note tracks the argument that machine learning outputs should often be represented as granules with coverage, specificity, and uncertainty rather than a single over-precise point estimate.
Granular Rules
From Fuzzy Rule-Based Models to Granular Models
This note explains how a fuzzy rule-based model can move from crisp consequents to interval, fuzzy, or probabilistic granules, changing both output semantics and uncertainty expression.
Knowledge Regularization
Knowledge Landmarks
This note follows how input-output landmark pairs and conditional fuzzy clustering become a knowledge-aware regularizer, with attention to local guidance, counterfactual meaning, and failure modes.
How these note pages are maintained
The writing stays in the Knowledge/ directory as Markdown, while the publication list lives in
tools/notes_manifest.json. The generator builds this overview page and the per-note HTML pages from that manifest.