Source-linked AI summary
Few-Shot Class-Incremental Learning from an Open-Set Perspective
Can Peng, Kun Zhao, Tianren Wang, Meng Li, Brian C. Lovell
TL;DR
FSCIL must learn continually arriving classes from scarce data without forgetting prior knowledge or becoming biased toward base classes. The paper reformulates the task and proposes ALICE, combining angular penalty loss with augmentation and balanced prototypes; it reports substantial improvements across three benchmark datasets, while acknowledging residual base-class bias in the one-shot setting.
Problem
FSCIL lacks evaluation that fully captures adaptation under extreme scarcity, balance between base and incremental classes, and retention of prior knowledge.
Method
ALICE uses angular penalty loss, augmentation, balanced prototype generation, nearest class mean, and cosine similarity for FSCIL.
Results
ALICE substantially outperforms state-of-the-art methods on CIFAR100, miniImageNet, and CUB200.
Takeaways & Limitations
Balanced nearest class mean classification with a generalizable feature extractor supports FSCIL without further prototype evolution in the reported benchmarks.
Takeaways & Limitations
In the one-shot setting, prediction bias toward base classes still exists despite outperforming CEC, motivating future work.
Abstract
from arXiv · showhide
The continual appearance of new objects in the visual world poses considerable challenges for current deep learning methods in real-world deployments. The challenge of new task learning is often exacerbated by the scarcity of data for the new categories due to rarity or cost. Here we explore the important task of Few-Shot Class-Incremental Learning (FSCIL) and its extreme data scarcity condition of one-shot. An ideal FSCIL model needs to perform well on all classes, regardless of their presentation order or paucity of data. It also needs to be robust to open-set real-world conditions and be easily adapted to the new tasks that always arise in the field. In this paper, we first reevaluate the current task setting and propose a more comprehensive and practical setting for the FSCIL task. Then, inspired by the similarity of the goals for FSCIL and modern face recognition systems, we propose our method -- Augmented Angular Loss Incremental Classification or ALICE. In ALICE, instead of the commonly used cross-entropy loss, we propose to use the angular penalty loss to obtain well-clustered features. As the obtained features not only need to be compactly clustered but also diverse enough to maintain generalization for future incremental classes, we further discuss how class augmentation, data augmentation, and data balancing affect classification performance. Experiments on benchmark datasets, including CIFAR100, miniImageNet, and CUB200, demonstrate the improved performance of ALICE over the state-of-the-art FSCIL methods.
1 Introduction
The paper frames FSCIL as learning new classes from scarce data while retaining prior knowledge, and proposes ALICE to improve open-set generalization and classification balance.
- FSCIL requires rapid adaptation to continually arriving tasks while retaining knowledge of previously learned tasks under severe data scarcity.
- Ideal FSCIL models should perform equally across classes and presentation orders while remaining robust in one-shot conditions.
- Current evaluation can miss prediction bias toward base classes because base and incremental classes differ in number and available data.
- The proposed method combines an open-set perspective with angular penalty loss, class and data augmentation, and balanced prototype construction.
- ALICE reports significant improvements over state-of-the-art methods on CIFAR100, miniImageNet, and CUB200.
2 Related Work
Prior FSCIL methods either update or freeze the backbone, but both approaches face practical trade-offs involving forgetting, overfitting, memory, data scarcity, and class imbalance.
- Backbone-updating FSCIL methods commonly use knowledge distillation, but may require stored exemplars or complex updates and struggle with one-shot scarcity.
- Freezing the backbone balances stability and plasticity by decoupling representation learning from classifier learning, reducing overfitting and catastrophic forgetting in representations.
- This paper focuses on feature distributions rather than only classifier design because feature quality is presented as foundational for robust classification.
- Angular-margin methods from face recognition enforce greater intra-class compactness and inter-class discrepancy, motivating their relevance to FSCIL.
- Existing metric-learning approaches often assume sufficient data, unlike FSCIL methods designed for limited incremental data and extreme scarcity.
3 Problem Formulation
The paper reformulates FSCIL evaluation around realistic incremental sessions, extreme scarcity, class-balance sensitivity, and dataset diversity, while motivating compact feature distributions for future classes.
- Task setting: FSCIL consists of a data-rich base task followed by multiple data-limited incremental tasks, with testing over all classes seen so far.
- Few-shot data: The proposed setup adds 1-shot evaluation because existing benchmarks rarely assess this extreme scarcity condition.
- Evaluation metric: Harmonic accuracy complements average accuracy by revealing prediction bias between base and incremental classes.
- Dataset diversity: Experiments span general datasets and fine-grained CUB200 to reduce performance benefits caused by high similarity between base and incremental classes.
- Feature design: ALICE uses angular penalty loss to produce compact intra-class clusters and wide inter-class separation, leaving room for future incremental classes.
4 Methodology
ALICE combines angular-penalty training, augmented data, and balanced prototype construction for FSCIL. Its design seeks compact, separated features that generalize to new classes while limiting prediction bias across base and incremental sessions.
- ALICE overview: ALICE trains the base-session feature extractor with angular penalty, class augmentation, data augmentation, and balanced data for incremental prototypes.The incremental classifier combines nearest class mean with cosine similarity.
- Angular similarity: Cosine similarity compares normalized features with normalized class weights, making each weight vector a category center and measuring angular class affinity.The fully connected bias is fixed to zero for this computation.
- Angular penalty: Angular penalty adds a cosine margin to cross-entropy so intra-class features cluster more tightly and inter-class features separate more widely.The experiments use scale factor s=30 and margin m=0.4.
- Augmented training: Class augmentation interpolates samples from different base classes to create auxiliary classes and expose the extractor to more varied representations.With C base classes, pairwise interpolation adds C × (C − 1)/2 auxiliary classes; λ is sampled from [0.4, 0.6].
- Augmented training: Two random transformations per image increase training diversity, after which the projection and augmented classification heads are discarded and the extractor is frozen.Freezing is intended to avoid overfitting and catastrophic forgetting during incremental sessions.
- Evaluation: Harmonic accuracy combines base-session and incremental-session average accuracy, exposing imbalance that class-wise average accuracy can hide.An ideal balanced classifier performs well on both components; poor harmonic accuracy indicates weak incremental learning despite strong base performance.
5 Experiments
ALICE is evaluated against state-of-the-art FSCIL methods across three datasets, using 5-shot and 1-shot settings, average and harmonic accuracy, confusion matrices, ablations, feature visualizations, and hyper-parameter studies. Results show strong performance, improved balance between base and incremental classes, and effects from angular loss, augmentation, and balanced prototype generation.
- Comparison with state-of-the-art methods: ALICE achieves the highest class-wise accuracy across sessions on CIFAR100 and miniImageNet under both 5-shot and 1-shot settings.It also shows higher harmonic accuracy than CEC across sessions on both datasets.
- Comparison with state-of-the-art methods: 54.1% and 47.5% average accuracy, with 50.6% and 26.5% harmonic accuracy, are achieved on CIFAR100 in the final 5-shot and 1-shot sessions.These exceed CEC by 6.0% and 2.7% in average accuracy, and by 19.3% and 13.5% in harmonic accuracy, respectively.
- Comparison with state-of-the-art methods: 55.7% and 48.6% average accuracy, with 50.9% and 27.1% harmonic accuracy, are achieved on miniImageNet in the final 5-shot and 1-shot sessions.These exceed CEC by 8.5% and 4.9% in average accuracy, and by 22.8% and 19.3% in harmonic accuracy, respectively.
- Comparison with state-of-the-art methods: On CUB200, class augmentation deteriorates performance because it obscures fine-grained class differences, so ALICE uses angular penalty and data augmentation without class augmentation.Under 1-shot, ALICE obtains similar average accuracy to CEC because both freeze the backbone after base-session training.
- Ablation study: Balanced prototype generation improves harmonic accuracy by using equal amounts of base and incremental data, alleviating prediction bias from data imbalance.The ablation combines angular penalty loss, two image transformations, class augmentation, and balanced prototype generation to assess their effects.
- Feature visualization: Angular penalty loss produces better latent-space clustering than cross-entropy, while projection layers and augmentation preserve class separation alongside diverse, transferable features.Figure 6 visualizes embeddings for the 60 CIFAR100 base classes and their average prototypes.
6 Conclusion
The paper reformulates FSCIL with a more practical setup and proposes ALICE, linking FSCIL to open-set recognition and emphasizing generalizable base-session features. Using balanced nearest class mean without further prototype evolution, ALICE reportedly outperforms state-of-the-art methods across all benchmark datasets.
- The paper proposes a more practical and comprehensive FSCIL setup.
- ALICE connects FSCIL with open-set tasks and uses base-session training to obtain generalizable features.
- ALICE uses balanced nearest class mean without further prototype evolution.
- ALICE achieves substantial improvements over state-of-the-art methods on all benchmark datasets.
1 Introduction
The supplementary experiments report average accuracy, harmonic accuracy, and performance dropping rate for FSCIL evaluation. Performance dropping rate measures the absolute accuracy decrease from base learning to the final incremental session.
- The supplementary material reports average accuracy and harmonic accuracy values in experiment tables.
- Performance dropping rate measures the absolute accuracy decrease between base learning and the last incremental session.
2 Detailed Results
The detailed results cover 5-shot and 1-shot FSCIL protocols across CIFAR100, miniImageNet, and CUB200. The experiments vary the number of incremental steps and classes per step by dataset and shot setting.
- 5-shot protocols: The 8-step 5-way 5-shot protocol is evaluated on CIFAR100 and miniImageNet.
- 5-shot protocols: The 10-step 10-way 5-shot protocol is evaluated on CUB200.
- 1-shot protocols: The 8-step 5-way 1-shot protocol is evaluated on CIFAR100 and miniImageNet.
- 1-shot protocols: The 10-step 10-way 1-shot protocol is evaluated on CUB200.