Research map / Continual learning / Class-incremental
DYSON: Dynamic Feature Space Self-Organization for Online Task-Free Class Incremental Learning
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 DYSON, a memory-free online task-free class-incremental learning method that computes a dynamically expanding neural-collapse simplex-ETF geometry (DNC) and aligns a projection layer's features to it, with a training-free classifier.
Problem
Online task-free class-incremental learning, where one-shot streaming samples arrive without task boundaries and the feature space must learn new classes while retaining old ones.
Approach
Instead of continuously learning the feature geometry from data, first compute an optimal simplex-ETF class-prototype geometry via a Dynamic Neural Collapse algorithm that expands when new classes arrive with bounded old-prototype drift, then align a trainable projection layer to it; the classifier is training-free nearest-prototype.
Key results
- On CIFAR-10 (ViT-S/8), DYSON with no buffer beats DSDM (1k buffer) by 7.1%, 7.9% and 8.9% under 1-step, 2-step and Gaussian protocols.
- On CIFAR-100 (ViT-S/8), DYSON outperforms DSDM by 16.1%, 14.8% and 15.0% under the same protocols.
- On CoRe50 (online CIL), DYSON achieves 70.7% Avg and 61.3% Last with no buffer, versus DSDM-5k at 66.3% Avg and 57.1% Last.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DYSON instead of DSDM with the same ViT-S/8 backbone | DYSON · CIFAR-10, 1-step | Last accuracy (%): 85.5 -> 92.6 | Helped |
| use DYSON instead of DSDM with the same ViT-S/8 backbone | DYSON · CIFAR-10, 2-step | Last accuracy (%): 85.6 -> 93.5 | Helped |
| use DYSON instead of DSDM with the same ViT-S/8 backbone | DYSON · CIFAR-10, Gaussian-step | Last accuracy (%): 84.9 -> 93.8 | Helped |
| use DYSON instead of DSDM with the same ViT-S/8 backbone | DYSON · CIFAR-100, 1-step | Last accuracy (%): 61.1 -> 77.7 | Helped |
| use DYSON instead of DSDM with the same ViT-S/8 backbone | DYSON · CIFAR-100, 5-step | Last accuracy (%): 60.8 -> 75.6 | Helped |
| use DYSON instead of DSDM with the same ViT-S/8 backbone | DYSON · CIFAR-100, Gaussian-step | Last accuracy (%): 61.4 -> 76.4 | Helped |
| use DYSON instead of DSDM | DYSON · CoRe50 | Avg accuracy (%): 43.9 -> 70.7 | Helped |
| use DYSON instead of DSDM | DYSON · CIFAR-100, 2-step | Avg accuracy (%): 63.3 -> 71.7 | Helped |
| use DYSON instead of DSDM | DYSON · CUB-200, 2-step | Avg accuracy (%): 55.2 -> 59.1 | Helped |
| remove prototype placeholders (M=0 instead of M=10) (ablation) | DYSON · CIFAR-100, 5-step | Last accuracy (%): 59.18 -> 56.80 | Hurt |
| use M=20 prototype placeholders instead of M=10 (ablation) | DYSON · CIFAR-100, 5-step | Last accuracy (%): 59.18 -> 58.24 | Hurt |
| use M=50 prototype placeholders instead of M=10 (ablation) | DYSON · CIFAR-100, 5-step | Last accuracy (%): 59.18 -> 59.01 | Hurt |
Surprising findings
- Memory-free DYSON outperforms methods that store 1k-6k exemplars on several benchmarks.
- Using a self-supervised DINO-ResNet50 backbone does not uniformly help: on CIFAR-100 1-step it improves, but 2-step drops (57.1 vs 59.2 with ResNet50).
Limitations
- Relies on pre-trained ImageNet backbones; new classes may already be seen by pretraining (authors partially address with DINO ablation).
- Only the projection layer is trained; feature extractor frozen, limiting adaptation to large domain shifts.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 CUB-200 CoRe50
Benchmarks: CIFAR-10 CIFAR-100 CUB-200 CoRe50
Metrics: last accuracy average accuracy
Baselines: ODDL SDP CoPE CN-DPM GMED FCA Ensemble DSDM L2P A-GEM MIR GSS
Builds on: neural collapse DSDM FCA
Closest papers by meaning
- Reducing Class-wise Confusion for Incremental Learning with Disentangled Manifolds (CVPR 2025 · similarity 0.85)
- Compress to One Point: Neural Collapse for Pre-Trained Model-Based Class-Incremental Learning (AAAI 2025 · similarity 0.84)
- Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning (CVPR 2025 · similarity 0.84)
- Order-Robust Class Incremental Learning: Graph-Driven Dynamic Similarity Grouping (CVPR 2025 · similarity 0.83)
- On Feature Learning in Structured State Space Models (NeurIPS 2024 · similarity 0.83)
- Expandable Subspace Ensemble for Pre-Trained Model-Based Class-Incremental Learning (CVPR 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.