Research map / Label-efficient and robust learning / Knowledge distillation
Flipped Classroom: Aligning Teacher Attention with Student in Generalized Category Discovery
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FlipClass, a teacher-student framework for generalized category discovery that aligns teacher attention with the student, evaluated on generic and fine-grained image benchmarks against XCon, PCAL, SPTNet, GCD, BaCon, and SimGCD.
Problem
Generalized category discovery (GCD) must classify both known and unseen classes without priors for new classes, and existing semi-supervised teacher-student approaches struggle with this.
Approach
Applies a teacher-student framework with representation alignment between teacher attention and student, plus an alpha-weighted alignment term; the paper reports that distribution-alignment strategies such as ReMixMatch-style alignment do not help when combined with its representation alignment.
Key results
- On Stanford Cars, FlipClass reaches 63.1 All / 81.7 Old / 53.8 New, versus SPTNet's 59.0 / 79.2 / 49.3.
- Average over six generic and fine-grained datasets is 64.6 for FlipClass versus 61.4 for SPTNet; on CIFAR-10/100 and ImageNet-100 average is 90.1 vs 88.0.
- On CUB, FlipClass reaches 71.3 All, versus SPTNet's 65.8.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use FlipClass instead of SPTNet | GCD classifier · Cifar10, Cifar100, ImageNet100, CUB, Stanford Cars, Aircraft | accuracy (All): 97.3 -> 98.5; 81.3 -> 85.2; 85.4 -> 86.7; 65.8 -> 71.3; 59.0 -> 63.1; 59.3 -> 59.3 | Mixed |
| use FlipClass instead of BaCon-S | long-tailed GCD · Cifar10-LT, Cifar100-LT, Places-LT | accuracy (All): 91.1 -> 72.8; 67.2 -> 60.1; 29.9 -> 26.6 | Hurt |
| add distribution alignment with alpha=0.01 instead of alpha=0 (ablation) | FlipClass · Stanford Cars | accuracy (All): 63.1 -> 52.9 | Hurt |
| add distribution alignment with alpha=0.05 instead of alpha=0 (ablation) | FlipClass · Stanford Cars | accuracy (All): 63.1 -> 41.8 | Hurt |
| add distribution alignment with alpha=0.1 instead of alpha=0 (ablation) | FlipClass · Stanford Cars | accuracy (All): 63.1 -> 32.9 | Hurt |
| add ReMixMatch Distribution Alignment to representation alignment (ablation) | FlipClass · Stanford Cars | accuracy (All): 63.1 -> 60.3 | Hurt |
| add Prior Distribution Alignment to representation alignment (ablation) | FlipClass · Stanford Cars | accuracy (All): 63.1 -> 58.1 | Hurt |
Surprising findings
- Adding ReMixMatch or prior distribution alignment on top of the representation alignment lowers Stanford Cars accuracy (60.3 and 58.1 versus 63.1).
Limitations
- Distribution-alignment strategies reduce performance when combined with the proposed method.
- Only visual and image benchmarks are evaluated in the provided text.
Experimental setup
Datasets: CUB Stanford Cars Aircraft Cifar10 Cifar100 ImageNet100 Cifar10-LT Cifar100-LT Places-LT
Benchmarks: CUB Stanford Cars Aircraft Cifar10 Cifar100 ImageNet100 Cifar10-LT Cifar100-LT Places-LT
Metrics: clustering accuracy (All/Old/New)
Baselines: XCon PCAL SPTNet GCD BaCon SimGCD ORCA OpenCon InfoSieve
Builds on: SimGCD ReMixMatch
Closest papers by meaning
- When Domain Generalization meets Generalized Category Discovery: An Adaptive Task-Arithmetic Driven Approach (CVPR 2025 · similarity 0.88)
- Happy: A Debiased Learning Framework for Continual Generalized Category Discovery (NeurIPS 2024 · similarity 0.87)
- HiLo: A Learning Framework for Generalized Category Discovery Robust to Domain Shifts (ICLR 2025 · similarity 0.87)
- SPTNet: An Efficient Alternative Framework for Generalized Category Discovery with Spatial Prompt Tuning (ICLR 2024 · similarity 0.87)
- Generalized Category Discovery via Reciprocal Learning and Class-Wise Distribution Regularization (ICML 2025 · similarity 0.86)
- Transfer and Alignment Network for Generalized Category Discovery (AAAI 2024 · similarity 0.86)
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.