Research map / Label-efficient and robust learning / Recognition and long-tail
Explaining Local, Global, and Higher-Order Interactions in Deep Learning
Samuel Lerman, Charles Venuto, Henry Kautz, Chenliang Xu
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 changed | Where | Measured | Effect |
|---|---|---|---|
| use T-NID instead of NID for 3-way interactions | T-NID · Synthetic functions 3-way | average AUC: 0.08 -> 0.76 | Improved |
| use T-NID instead of NID for 4-way interactions | T-NID · Synthetic functions 4-way | average AUC: 0.75 -> 0.78 | Improved |
| use T-NID instead of NID for 5-way interactions | T-NID · Synthetic functions 5-way | average AUC: 0.92 -> 0.97 | Improved |
| use T-NID instead of GLIDER for pairwise interactions | T-NID · Synthetic functions pairwise | average AUC: 0.892 -> 0.975 | Improved |
| use T-NID instead of NID for pairwise interactions | T-NID · Synthetic functions pairwise | average AUC: 0.92 -> 0.975 | Improved |
| add main-effects network (MLP-M) and sparsity regularization to T-NID (ablation) | T-NID · Synthetic functions n-way | average AUC | No clear effect |
| add MLP-M for 4-way interactions (T-NID vs NID, both MLP-M) (ablation) | T-NID · Synthetic functions 4-way | average AUC: 0.723 vs 0.713 | No clear effect |
| use Taylor-CAM instead of Grad-CAM* (Ques 1) | Taylor-CAM · Sort-Of-CLEVR | top-interaction accuracy (%): 35% -> 90% | Improved |
| use Taylor-CAM instead of GLIDER (Ques 2) | Taylor-CAM · Sort-Of-CLEVR | top-interaction accuracy (%): 35% -> 55% | Improved |
| use Taylor-CAM instead of Grad-CAM* (Ques 2) | Taylor-CAM · Sort-Of-CLEVR | top-interaction accuracy (%): 50% -> 55% | Improved |
| use Taylor-CAM instead of Grad-CAM for question guessing (Ques 2) | Taylor-CAM · User study question accuracy | accuracy (%): 14% -> 55% | Improved |
| use Taylor-CAM instead of Grad-CAM for question guessing (Ques 1) | Taylor-CAM · User study question accuracy | accuracy (%): 44% -> 76% | Improved |
Closest papers by meaning
- Defining and extracting generalizable interaction primitives from DNNs (ICLR 2024 · similarity 0.86)
- On the Joint Interaction of Models, Data, and Features (ICLR 2024 · similarity 0.85)
- Interaction Asymmetry: A General Principle for Learning Composable Abstractions (ICLR 2025 · similarity 0.85)
- Towards the Dynamics of a DNN Learning Symbolic Interactions (NeurIPS 2024 · similarity 0.85)
- SPEX: Scaling Feature Interaction Explanations for LLMs (ICML 2025 · similarity 0.85)
- Monitoring Primitive Interactions During the Training of DNNs (AAAI 2025 · similarity 0.85)
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.
