Research map / Label-efficient and robust learning / Recognition and long-tail
Transferability and Hardness of Supervised Classification Tasks
Anh T. Tran, Cuong V. Nguyen, Tal Hassner
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 |
|---|---|---|---|
| transfer a face recognition network (lSVM on embeddings) instead of training a dedicated network, for highly transferable attributes | ResNet101 face recognition embedding · CelebA Male | accuracy: 0.985 -> 0.992 | Improved |
| transfer a face recognition network instead of dedicated training, for weakly transferable attributes | ResNet101 face recognition embedding · CelebA Mouth Open | accuracy: 0.943 -> 0.901 | Hurt |
| transfer a face recognition network instead of dedicated training, averaged over 40 attributes | ResNet101 face recognition embedding · CelebA (average over 40 attributes) | accuracy: 0.911 -> 0.902 | Hurt |
| transfer instead of dedicated training for attributes young and big nose | face recognition transfer · CelebA | accuracy | Improved |
| transfer from recognition vs best published result DMTL 2018 | face recognition transfer + lSVM · CelebA | accuracy: 0.926 -> 0.902 | Hurt |
| transfer from face recognition with far less training data | face recognition transfer + lSVM · CelebA Male and Double Chin | accuracy | Improved |
| higher conditional entropy between source and target labels | CE estimate · CelebA, AwA2, CUB attribute tasks | target test error | Hurt |
| higher conditional entropy of labels (estimated task hardness) | CE H(Z|C) · CelebA, AwA2, CUB | test error of dedicated networks: correlation significant with p<0.001 | Hurt |
| higher estimated hardness on multi-class CelebA aggregates | hardness estimate · CelebA multi-class | accuracy: Straight/Wavy/Other hardness 1.040, accuracy 0.713; Goatee/Beard/None hardness 0.557, accuracy 0.937 | Hurt |
| lower transferability (larger CE) from identity to attribute | face recognition transfer · CelebA | transferred minus dedicated accuracy: Male 0.992 vs 0.985; Smiling 0.909 vs 0.933 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- MOON 2016 — MOON: A Mixed Objective Optimization Network for the Recognition of Facial Attributes (ECCV 2016) · better on 1 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
- Understanding the Transferability of Representations via Task-Relatedness (NeurIPS 2024 · similarity 0.85)
- Frustratingly Easy Transferability Estimation (ICML 2022 · similarity 0.85)
- Distribution Matching for Multi-Task Learning of Classification Tasks: A Large-Scale Study on Faces & Beyond (AAAI 2024 · similarity 0.85)
- Exploring Transferability of Self-Supervised Learning by Task Conflict Calibration (AAAI 2026 · similarity 0.85)
- Labels, Information, and Computation: Efficient Learning Using Sufficient Labels (JMLR 2023 · similarity 0.85)
- Towards the Fundamental Limits of Knowledge Transfer over Finite Domains (ICLR 2024 · 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.
