Research map / Label-efficient and robust learning / Recognition and long-tail
SeA: Semantic Adversarial Augmentation for Last Layer Features from Unsupervised Representation Learning
Qi Qian, Yuanhong Xu, Juhua Hu
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 |
|---|---|---|---|
| add semantic adversarial augmentation to fixed-feature linear training (average over 11 tasks) (ablation) | linear classifier on S+M+B+C features · 11 downstream tasks average | accuracy (%): 81.1 -> 83.1 | Improved |
| increase augmentation step size from 0 to an appropriate value (ablation) | SeA · CIFAR-100 | accuracy (%): 81.1 -> 82.9 | Improved |
| random augmentation direction in the batch subspace instead of the semantic adversarial direction (ablation) | SeA · CIFAR-100 | accuracy (%) | No clear effect |
| use raw gradient adversarial direction (Adv) instead of projected semantic direction (ablation) | SeA · CIFAR-100 | accuracy (%) | Hurt |
| feature-space SeA instead of image-space augmentation in linear probing (ablation) | single MoCo model · CIFAR-100 | accuracy (%): 77.8 -> 79.0 | Improved |
| concatenate features from four pre-trained models instead of one model (ablation) | SeA · 11 downstream tasks | accuracy (%): CIFAR-100 78.5 (C) -> 82.9 (S+M+B+C) | Improved |
| use fixed deep features with SeA instead of fine-tuning (Aircraft, Cars) | SeA · Aircraft | accuracy (%): 89.0 -> 67.4 | Hurt |
| use fixed deep features with SeA instead of fine-tuning (DTD) | SeA · DTD | accuracy (%): 78.6 -> 79.8 | Improved |
| use SeA instead of ISDA feature-space augmentation | feature ensemble · 11 tasks average | accuracy (%): 81.2 -> 83.1 | Improved |
| train on fixed features with SeA instead of fine-tuning (cost) | SeA · CIFAR-100 running time | time (s): 14,383 -> 40 | Improved |
Closest papers by meaning
- Implicit Semantic Data Augmentation for Deep Networks (NeurIPS 2019 · similarity 0.86)
- Adversarial Auto-Augment with Label Preservation: A Representation Learning Principle Guided Approach (NeurIPS 2022 · similarity 0.86)
- Regularizing Deep Networks with Semantic Data Augmentation (TPAMI 2021 · similarity 0.85)
- Generalizing to Unseen Domains via Adversarial Data Augmentation (NeurIPS 2018 · similarity 0.85)
- Classes Matter: A Fine-grained Adversarial Approach to Cross-domain Semantic Segmentation (ECCV 2020 · similarity 0.85)
- Minimal Semantic Sufficiency Meets Unsupervised Domain Generalization (NeurIPS 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.
