Research map / Label-efficient and robust learning / Recognition and long-tail
Defining and extracting generalizable interaction primitives from DNNs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a loss that extracts AND-OR interaction primitives shared across multiple DNNs trained on the same task, improving their cross-model generalization power over traditional sparsity-only extraction.
Problem
Interactions between input variables proposed as primitive inference patterns of DNNs are not guaranteed to be generalizable; interactions extracted from one DNN may not be shared by others, making them unfaithful as concepts.
Approach
Learn decomposition parameters jointly across multiple DNNs using a rowmax (l-infinity over models) penalty on salient interactions plus an l1 penalty on all interactions, with shared and model-specific decomposition terms and bounded error terms, so that interactions salient for any DNN are encouraged to appear in all DNNs.
Key results
- Traditional extraction on BERT models yields two sets of interactions with only 21% overlap across initializations.
- The proposed method extracts interactions with higher generalization power than traditional interactions across BERT-base/large, LLaMA/OPT, and ResNet-20/VGG-16 pairs.
- Low-order interactions are more likely to be shared across DNNs than high-order interactions.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| extract interactions jointly across different DNNs instead of using traditional interaction extraction | AND-OR interaction extraction · SST-2, sentiment classification | generalization power | Helped |
| extract interactions jointly across different DNNs instead of using traditional interaction extraction | AND-OR interaction extraction · Three tasks: SST-2, SQuAD, and MNIST | shared interaction ratio | Helped |
| use the proposed generalizable interaction extraction method | Interaction primitives · Three tasks: SST-2, SQuAD, and MNIST | sparsity | Hurt |
Surprising findings
- Interactions extracted from ResNet-20 and VGG-16, which have fully different architectures, show high similarity, suggesting a shared set of ultimate interactions for a task.
Limitations
- Computational cost is exponential (2^n) in the number of input variables, requiring restriction to a few selected variables or patches.
- The method reduces sparsity somewhat relative to the traditional method.
- Interactions are defined via masking with baseline values, so the generalization claims depend on the masking choice.
Experimental setup
Datasets: SST-2 SQuAD MNIST
Benchmarks: SST-2 SQuAD MNIST
Metrics: generalization power s_and and s_or sparsity of interaction strengths matching error ratio of shared interactions by order
Baselines: Traditional interaction extraction Harsanyi interactions
Builds on: Harsanyi dividend universal matching theorem AND-OR interactions
Closest papers by meaning
- Layerwise Change of Knowledge in Neural Networks (ICML 2024 · similarity 0.90)
- Monitoring Primitive Interactions During the Training of DNNs (AAAI 2025 · similarity 0.90)
- Towards the Dynamics of a DNN Learning Symbolic Interactions (NeurIPS 2024 · similarity 0.89)
- Explaining Generalization Power of a DNN Using Interactive Concepts (AAAI 2024 · similarity 0.88)
- Beyond Topological Self-Explainable GNNs: A Formal Explainability Perspective (ICML 2025 · similarity 0.85)
- Explainable Neural Networks with Guarantee: A Sparse Estimation Approach (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.