Research map / Label-efficient and robust learning / Recognition and long-tail
An Information-Geometric Distance on the Space of Tasks
Yansong Gao, Pratik Chaudhari
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 coupled task transport and synchronized weight updates to estimate task distance | 8-layer CNN task transfer · CIFAR-10 and CIFAR-100 task pairs | Mantel r with fine-tuning distance: 0.03 -> 0.428 | Improved |
| Use coupled task transport and synchronized weight updates to estimate task distance | 8-layer CNN task transfer · CIFAR-100 subset tasks | Mantel r with fine-tuning distance: 0.07 -> 0.14 | Improved |
| Use coupled task transport and synchronized weight updates to estimate task distance | WRN-16-4 task transfer · Deep Fashion task pairs | Mantel r with fine-tuning distance: 0.04 -> 0.37 | Improved |
| Synchronize task transport with weight updates rather than use uncoupled interpolation (ablation) | CIFAR-100 subset task transfer · CIFAR-100 subset tasks | Distance estimate relative to fine-tuning: 0.12 -> 0.14 | Improved |
| Replace the smaller CNN with a higher-capacity WRN-16-4 (ablation) | CIFAR-100 subset task transfer · CIFAR-100 subset tasks | Coupled transfer distance | Improved |
| Use coupled OT task interpolation during transfer instead of standard fine-tuning | CIFAR-100 subset task transfer · CIFAR-100 subset tasks | Validation loss | Improved |
| Use coupled OT task interpolation during transfer instead of standard fine-tuning | CIFAR-100 subset task transfer · CIFAR-100 subset tasks | Validation accuracy: 0.4%--2.5% | Mixed |
| Optimize transport coupling jointly with the weight trajectory rather than compute the coupling independently (ablation) | Coupled transfer · Task transfer | Weight trajectory distance | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Task2Vec — Task2Vec: Task Embedding for Meta-Learning (ICCV 2019) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Geometric Dataset Distances via Optimal Transport (NeurIPS 2020 · similarity 0.88)
- Geometry and Stability of Supervised Learning Problems (JMLR 2025 · similarity 0.86)
- Task-Induced Riemannian Metrics for Vision Transformer Feature Spaces (NeurIPS 2026 · similarity 0.86)
- Online Learning in Variable Feature Spaces under Incomplete Supervision (AAAI 2021 · similarity 0.85)
- Learning Distances from Data with Normalizing Flows and Score Matching (ICML 2025 · similarity 0.85)
- Stabilizing Sample Similarity in Representation via Mitigating Random Consistency (ICML 2025 · similarity 0.84)
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.
