Research map / Probabilistic and scientific modeling / Molecules and proteins
Neural Interaction Transparency (NIT): Disentangling Learned Interactions for Improved Interpretability
Michael Tsang, Hanpeng Liu, Sanjay Purushotham, Pavankumar Murali, Yan Liu
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 |
|---|---|---|---|
| raise max interaction order K from 2 to 4 (ablation) | NIT · Bike Sharing | RMSE: 0.31 ± 0.013 -> 0.240 ± 0.0097 | Improved |
| raise max interaction order K from 2 to 4 (ablation) | NIT · Cal Housing | RMSE: 0.448 ± 0.0080 -> 0.43 ± 0.013 | Improved |
| raise max interaction order K from 2 to 6 (ablation) | NIT · MIMIC-III | AUC: 0.76 ± 0.011 -> 0.77 ± 0.011 | No clear effect |
| raise max interaction order K from 10 to 20 (ablation) | NIT · CIFAR-10 binary | AUC: 0.849 ± 0.0049 -> 0.860 ± 0.0034 | Improved |
| allow pairwise interactions (NIT K=2) instead of univariate GAM | GAM · Bike Sharing | RMSE: 0.55 ± 0.016 -> 0.31 ± 0.013 | Improved |
| disentangle interactions with NIT instead of plain L1/L2 regularized MLP | feedforward network · synthetic x1x2 + x3x4 | entangled neurons in second layer | Improved |
| disentangle through all layers without network blocks (ablation) | NIT · Cal Housing | RMSE: 0.43 RMSE at max order 3 | No clear effect |
| disentangle through all layers without network blocks (ablation) | NIT · Bike Sharing | RMSE: 0.26 RMSE at max order 6 | No clear effect |
Closest papers by meaning
- Defining and extracting generalizable interaction primitives from DNNs (ICLR 2024 · similarity 0.86)
- Interaction Asymmetry: A General Principle for Learning Composable Abstractions (ICLR 2025 · similarity 0.86)
- Layerwise Change of Knowledge in Neural Networks (ICML 2024 · similarity 0.84)
- Monitoring Primitive Interactions During the Training of DNNs (AAAI 2025 · similarity 0.84)
- Disentangling Representations through Multi-task Learning (ICLR 2025 · similarity 0.84)
- SPEX: Scaling Feature Interaction Explanations for LLMs (ICML 2025 · similarity 0.83)
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.
