AI Paper Atlas

Research map / Graph learning / Graph algorithms and clustering

Neuron Dependency Graphs: A Causal Abstraction of Neural Networks

Yaojie Hu, Jin Tian

ICML 2022 · Graph learning › Graph algorithms and clustering · approach: Continuous DAG structure learning

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Use random-initialized (untrained) model instead of trained model for NDG extraction (ablation)NDG edges · AllNLI, CUB200, Devignedge accuracy on test set: test accuracy ~50% vs ~99% for trainedHurt
Feed random inputs to trained model instead of real inputs (ablation)NDG edges · MNIST, SST2, AllNLI, CUB200, Devignedge test accuracy: e.g. CUB200 91.80 vs 97.15 original; AllNLI 99.55 vs 99.58Mixed
Build inter-model edges between two independently trained models on same input (ablation)NDG · MNIST even, SST2, AllNLI, Devignedge count and test accuracy: 0 edges on MNIST and CUB200; 35354 edges on SST2 with 92.29 test accuracyMixed
Same model with two disjoint input halves (sanity check) (ablation)NDG · all datasetsnumber of inter-model edges: 0 edges on all datasetsNo clear effect
Increase threshold alpha from low to high values (ablation)NDG edges · all datasetstraining and test edge accuracy: accuracy generally increases with alphaImproved
Interchange intervention on NDG-aligned neurons (Algorithm 1)neural network predictions · MNIST, MNIST even, SST2, AllNLI, CUB200, Devignaligned percentage on test set: 88.39% to 100% alignedImproved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy