Survey
Begin with the taxonomy so that later methods can be compared as different interfaces for the integration of prior knowledge.
Reading Map
Read the methods in terms of optimization interface, then move to code
This page gives a compact reading order. It is meant to clarify what to read first, why that order may be useful, and which toy reproduction becomes more interpretable after each step.
Logic-Net
Then examine how logic rules are converted into a teacher distribution and later written back into parameters through distillation.
Semantic Loss
Then revisit Semantic Loss until the distinction between class probabilities and satisfying-world mass becomes operationally clear.
DL2
Now study DL2 as a broader declarative-constraint route, comparing surrogate losses, feasible sets, and query-time optimization.
Granular / Landmarks
Then extend from exact and declarative constraints toward the granular and knowledge-landmark papers, while exploring knowledge-guided feature selection.
Suggested reproduction order
Use these toys not only to reproduce mechanisms, but to trace where knowledge actually enters the learning process: teacher distribution, satisfying-mass loss, or softer global regularization.
Long-form roadmap document
The full roadmap remains in the repository and includes screenshots, note relationships, and code-reading guidance that go beyond the short overview on this page.