研究项目Python研究

neusym-bridge

从神经网络表征中抽出符号因果结构。

神经符号桥接实验：在多个神经网络的共同表征中检测结构（CKA、Procrustes），并抽取可用符号规则表达的因果结构。

Python最后更新 2026年4月12日

GitHub 仓库 ↗

下载

这个项目还没有正式发布。

可以先去仓库看看源码，有新版本时这里会自动更新。

文档

neusym-bridge

Neurosymbolic bridge: extracting symbolic causal structure from neural network common representations.

Project Structure

neusym-bridge/
├── configs/                    # Experiment configurations
├── docs/results/               # Experiment reports and archived results
├── rules/                      # Relatum rule files
├── scripts/                    # Phase orchestration scripts
│   ├── run_phase0.py           # Data generation + model training
│   ├── run_phase1.py           # Common structure detection (CKA, Procrustes)
│   ├── run_phase2.py           # Physics correspondence (SVCCA, SINDy)
│   └── run_phase3.py           # Collapse mechanism verification
├── src/neusym_bridge/
│   ├── data/                   # PDE solvers and data generation
│   ├── models/                 # CNN world model + training
│   ├── analysis/               # Representation analysis + structure extraction
│   └── relatum/                # Probabilistic collapse mechanism
└── tests/                      # 46 tests

Validation Phases

Phase	Status	Key Result
0: Environment	PASS	3 models converge (ratio=1.006)
1: Common Structure	PASS 4/5	CKA=0.944 across 3 CNN models
2: Physics Correspondence	Conditional	SVCCA=0.999, 3 causal directions found
3: Collapse Mechanism	PASS 6/6	Collapse, retraction, active query verified