Research map / Continual learning / Class-incremental
AnchorInv: Few-Shot Class-Incremental Learning of Physiological Signals via Feature Space-Guided Inversion
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 AnchorInv, a few-shot class-incremental learning method for physiological signals that inverts synthetic replay samples from feature-space anchor points, evaluated on BCI-IV 2a, NHIE, and GRABMyo.
Problem
Few-shot class-incremental learning for physiological time series, where few base classes give weak generalizable backbones and storing raw patient data raises privacy concerns, causing catastrophic forgetting during incremental finetuning.
Approach
Select anchor points summarizing each session's feature distribution and invert synthetic replay samples whose embeddings match these anchors (feature-space-guided model inversion), then finetune the backbone with these replay samples plus few-shot new-class data, instead of storing raw exemplars or guiding inversion through the label space.
Key results
- On BCI 1-way-10-shot, AnchorInv reaches 37.04 ± 1.24 macro-F1 (all classes) in session 2 versus 32.69 for TEEN and 32.44 for DeepInv, a 2.19-point gain over the best baseline.
- On NHIE, AnchorInv reaches 66.03 ± 2.73 macro-F1 in session 2 versus 62.79 for DeepInv, a 3.24-point gain over the best baseline.
- On GRABMyo, AnchorInv achieves 60.78 ± 0.66 macro-F1 in session 6 versus 60.15 for ProtoNet, with a smaller margin when more base classes are available.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AnchorInv instead of the strongest reported baseline | FSCIL with EEG-Conformer · BCI-IV 2a, Session 2 | Macro-F1 (All Classes): 34.85 ± 1.53 -> 37.04 ± 1.24 | Helped |
| Use AnchorInv instead of the strongest reported baseline | FSCIL with EEG-Conformer · NHIE, Session 2 | Macro-F1 (All Classes): 62.79 ± 2.89 -> 66.03 ± 2.73 | Helped |
| Use AnchorInv instead of the strongest reported baseline | FSCIL with EEG-Conformer · GRABMyo, Session 6 | Macro-F1 (All Classes): 60.15 ± 0.75 -> 60.78 ± 0.66 | Helped |
| Vary base-session anchor selection among closest-to-prototype, k-means, and random-closest strategies instead of random sampling (ablation) | AnchorInv · BCI-IV 2a, Session 2 | Macro-F1 (All Classes): 37.04±1.24; 38.65 ± 1.26; 36.97±1.28; 38.36±1.34; 38.55±1.30; 38.05±1.61 | Mixed |
| Vary base-session anchor selection among closest-to-prototype, k-means, and random-closest strategies instead of random sampling (ablation) | AnchorInv · NHIE, Session 2 | Macro-F1 (All Classes): 66.03±2.73; 62.63±2.38; 66.05±2.90; 63.07±2.39; 64.03±2.23; 64.21±2.33 | Mixed |
| Vary base-session anchor selection among closest-to-prototype, k-means, and random-closest strategies instead of random sampling (ablation) | AnchorInv · GRABMyo, Session 6 | Macro-F1 (All Classes): 60.78±0.66; 60.67±0.70; 60.70±0.68; 60.67±0.71; 60.76±0.57; 60.68±0.71 | No clear effect |
| Increase the number of base-session anchor points and corresponding inverted samples (ablation) | AnchorInv · BCI-IV 2a, last session | Macro-F1 | Helped |
| Increase incremental training shots per class (ablation) | AnchorInv · BCI-IV 2a, last session | Macro-F1 | Helped |
Surprising findings
- Generalization of a backbone trained on few base classes to unseen classes decreases as base classes are reduced, even though random chance performance rises.
- Real and inverted EEG samples with nearly identical feature vectors look visually distinct, with inverted samples dominated by high-frequency fluctuations.
Limitations
- Results are sensitive to the choice of few-shot incremental samples, requiring repeated trials to estimate performance.
- Gains over baselines are smaller on GRABMyo, where more base classes are available.
- Anchor selection strategy has no clearly dominant option across datasets.
Experimental setup
Datasets: BCI-IV 2a NHIE GRABMyo
Benchmarks: BCI-IV 2a NHIE GRABMyo
Metrics: Macro-F1 (All Classes) Macro-F1 (Base Classes) Macro-F1 (Incremental Classes)
Baselines: Finetune NC-FSCIL ProtoNet TEEN DeepDream DeepInv
Builds on: DeepInv DeepDream EEG-Conformer ProtoNet MAPIC
Closest papers by meaning
- Model Inversion with Layer-Specific Modeling and Alignment for Data-Free Continual Learning (NeurIPS 2025 · similarity 0.85)
- Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning (CVPR 2025 · similarity 0.84)
- Pre-trained Vision and Language Transformers Are Few-Shot Incremental Learners (CVPR 2024 · similarity 0.83)
- Multi-layer Rehearsal Feature Augmentation for Class-Incremental Learning (ICML 2024 · similarity 0.83)
- Compress to One Point: Neural Collapse for Pre-Trained Model-Based Class-Incremental Learning (AAAI 2025 · similarity 0.82)
- NICE: Neurogenesis Inspired Contextual Encoding for Replay-free Class Incremental Learning (CVPR 2024 · 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.