Research map / Label-efficient and robust learning / Recognition and long-tail
Adaptive wavelet distillation from neural networks through interpretations
Wooseok Ha, Chandan Singh, Francois Lanusse, Srigokul Upadhyayula, Bin Yu
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 |
|---|---|---|---|
| distill LSTM into adaptive wavelet instead of fixed DB5 wavelet | wavelet model on LSTM · CME molecular partner prediction | R^2: 0.197 -> 0.262 | Improved |
| AWD wavelet model versus the original LSTM | LSTM · CME molecular partner prediction | R^2: 0.237 -> 0.262 | Improved |
| use adaptive wavelet instead of DB5 for compression | wavelet model · CME molecular partner prediction | compression factor: 0.704 -> 0.574 | Improved |
| replace LSTM with AWD wavelet model at test time | LSTM · CME molecular partner prediction | computation time: 0.0449s -> 0.0002s | Improved |
| AWD wavelet filters instead of DB5 wavelet for peak finding | peak-counting predictor · Cosmological parameter prediction | RMSE x 10^-2: 1.569 -> 1.029 | Improved |
| AWD filters instead of Laplace filter | peak-counting predictor · Cosmological parameter prediction | RMSE x 10^-2: 1.369 -> 1.029 | Improved |
| AWD filters instead of Roberts-Cross filter | peak-counting predictor · Cosmological parameter prediction | RMSE x 10^-2: 1.259 -> 1.029 | Improved |
| AWD interpretable model instead of the full ResNet | ResNet18 · Cosmological parameter prediction | RMSE x 10^-2: 1.156 -> 1.029 | Improved |
| AWD instead of DB5 wavelet for compression | wavelet model · Cosmological parameter prediction | compression rate: 0.620 -> 0.610 | Improved |
| vary interpretation penalty gamma (too small or too large) (ablation) | AWD · Simulated DB5 recovery | distance to groundtruth wavelet | Mixed |
| initialize lowpass filter at Sym 5 versus DB5+noise, varying sparsity penalty lambda (ablation) | AWD · Simulated DB5 recovery | distance to groundtruth wavelet | Mixed |
| initialize lowpass filter at Coif 2, varying sparsity penalty lambda (ablation) | AWD · Simulated DB5 recovery | distance to groundtruth wavelet | Mixed |
Closest papers by meaning
- One Wave To Explain Them All: A Unifying Perspective On Feature Attribution (ICML 2025 · similarity 0.81)
- Large Scale Dataset Distillation with Domain Shift (ICML 2024 · similarity 0.81)
- Multiscale Attention Wavelet Neural Operator for Capturing Steep Trajectories in Biochemical Systems (AAAI 2024 · similarity 0.81)
- Enhancing Foundation Models for Time Series Forecasting via Wavelet-based Tokenization (ICML 2025 · similarity 0.81)
- Wavelet Flow Matching for Multi-Scale Physics Emulation (NeurIPS 2026 · similarity 0.81)
- Debiased Distillation for Consistency Regularization (AAAI 2025 · similarity 0.81)
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.
