Research map / Learning theory and algorithms / Prediction and decision losses
HyDRA: Hypergradient Data Relevance Analysis for Interpreting Deep Neural Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 HyDRA instead of Influence Function to rank and remove noisy training samples | Label-noise filtering · MNIST, 80% label noise | accuracy (%): 90.07 -> 98.31 | Helped |
| Use HyDRA instead of Influence Function to rank and remove noisy training samples | Label-noise filtering · Fashion-MNIST, 80% label noise | accuracy (%): 63.20 -> 86.72 | Helped |
| Use HyDRA instead of Influence Function to rank and remove noisy training samples | Label-noise filtering · CIFAR-10, 80% label noise | accuracy (%): 42.12 -> 72.01 | Helped |
| Use HyDRA instead of Influence Function to rank and remove noisy training samples | Label-noise filtering · MNIST, 50% label noise | accuracy (%): 93.83 -> 98.31 | Helped |
| Use HyDRA instead of Influence Function to rank and remove noisy training samples | Label-noise filtering · Fashion-MNIST, 50% label noise | accuracy (%): 76.70 -> 87.28 | Helped |
| Use HyDRA instead of Influence Function to rank and remove noisy training samples | Label-noise filtering · CIFAR-10, 50% label noise | accuracy (%): 58.34 -> 75.06 | Helped |
| Filter noisy data using HyDRA rather than keeping all training data | Label-noise filtering · Fashion-MNIST, 50% label noise | accuracy (%): 87.44 -> 87.28 | Hurt |
| Use HyDRA's Hessian-free approximation instead of Influence Function's local final-parameter estimate | Data-contribution estimation · Fashion-MNIST | sign error rate: 0.130 -> 0.014 | Helped |
| Use HyDRA's Hessian-free approximation instead of Influence Function's local final-parameter estimate | Data-contribution estimation · Fashion-MNIST | Spearman rank correlation: 0.820 -> 0.986 | Helped |
| Track data influence across the optimization trajectory rather than only around final parameters | Data-contribution estimation · Fashion-MNIST | contribution-estimation error: 11.6% to 16.6% | Helped |
| Remove the Hessian from the trajectory hypergradient calculation (ablation) | HyDRA · Fashion-MNIST, 500 tracked samples | average L2 approximation error: 0.74 | Mixed |
| Remove Hessian-vector products from contribution tracking (ablation) | HyDRA · DenseNet-40, 20000 tracked samples, 1 epoch | wall-clock time: about 49 hours -> about 3 minutes; 971 folds | Helped |
Closest papers by meaning
- Diving into the shallows: a computational perspective on large-scale shallow learning (NeurIPS 2017 · similarity 0.84)
- Outlier Gradient Analysis: Efficiently Identifying Detrimental Training Samples for Deep Learning Models (ICML 2025 · similarity 0.83)
- Domain2Vec: Vectorizing Datasets to Find the Optimal Data Mixture without Training (ICML 2025 · similarity 0.83)
- Explainable Neural Networks with Guarantee: A Sparse Estimation Approach (AAAI 2025 · similarity 0.83)
- Neural Collapse in Cumulative Link Models for Ordinal Regression: An Analysis with Unconstrained Feature Model (NeurIPS 2025 · similarity 0.83)
- Neural Collapse under Gradient Flow on Shallow ReLU Networks for Orthogonally Separable Data (NeurIPS 2025 · similarity 0.82)
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.