Source-linked AI summary
Self-Promoted Prototype Refinement for Few-Shot Class-Incremental Learning
Kai Zhu, Yang Cao, Wei Zhai, Jie Cheng, Zheng-Jun Zha
TL;DR
Few-shot class-incremental learning must learn new classes from few samples without forgetting old classes, despite limited supervision for representation and prototype updates. The paper proposes an incremental prototype learning scheme using random episode selection and self-promoted prototype refinement with dynamic relation projection. It reports margins of 13%, 17%, and 11% over state-of-the-art methods on three benchmark datasets.
Problem
Few-shot class-incremental learning must recognize new classes from few samples while retaining old-class recognition, but limited supervision makes representation optimization and prototype reorganization difficult.
Method
The scheme combines random episode selection for extensible representations with self-promoted prototype refinement guided by relations between new samples and old prototypes.
Results
The method leads state-of-the-art methods by 13%, 17%, and 11% on CIFAR-100, MiniImageNet, and CUB200, respectively.
Takeaways & Limitations
The proposed scheme incorporates few-shot incremental classes into recognition while improving performance and adaptability relative to state-of-the-art methods.
Abstract
from arXiv · showhide
Few-shot class-incremental learning is to recognize the new classes given few samples and not forget the old classes. It is a challenging task since representation optimization and prototype reorganization can only be achieved under little supervision. To address this problem, we propose a novel incremental prototype learning scheme. Our scheme consists of a random episode selection strategy that adapts the feature representation to various generated incremental episodes to enhance the corresponding extensibility, and a self-promoted prototype refinement mechanism which strengthens the expression ability of the new classes by explicitly considering the dependencies among different classes. Particularly, a dynamic relation projection module is proposed to calculate the relation matrix in a shared embedding space and leverage it as the factor for bootstrapping the update of prototypes. Extensive experiments on three benchmark datasets demonstrate the above-par incremental performance, outperforming state-of-the-art methods by a margin of 13%, 17% and 11%, respectively.
1. Introduction
Few-shot class-incremental learning must incorporate new classes from few samples while retaining old-class recognition. The proposed scheme combines random episode selection with self-promoted prototype refinement and reports leading benchmark performance.
- Motivation: Few-shot class-incremental learning addresses adding atypical new classes from few samples while preserving recognition of old classes.Representation optimization and prototype reorganization are difficult under limited supervision.
- Motivation: Existing incremental methods can suffer performance drops because compact representations lack extensibility and few new samples provide insufficient supervision for expansion.The paper attributes this limitation to the mismatch between standard CIL representations and FSCIL requirements.
- Proposed Method: The incremental prototype learning scheme uses random episode selection to adapt features to varied simulated incremental processes.Randomly generated episodes force gradients to accommodate different incremental scenarios.
- Proposed Method: Self-promoted prototype refinement updates existing prototypes using relations between new-class representations and old-class prototypes.The mechanism aims to strengthen new-class expression while retaining relational characteristics among old classes.
- Results: 13%, 17%, and 11% are the reported margins over state-of-the-art methods on CIFAR-100, MiniImageNet, and CUB200, respectively.The comparison covers three benchmark datasets.
2. Related Work
Prior work addresses continual recognition, few-shot adaptation, and class dependencies through exemplars, distillation, topology preservation, and learned generators. This paper instead emphasizes extensible representations and explicit inter-class dependencies for FSCIL.
- Class-Incremental Learning: Class-incremental methods preserve old knowledge using exemplars, distillation, classifier correction, and prototype-based classification.The related work includes iCARL, NCM, BiC, WA, and related approaches.
- FSCIL: TOPIC addresses FSCIL with a neural gas network that constrains feature-space topology under insufficient incremental samples.The paper follows TOPIC’s FSCIL settings but uses a different update strategy.
- Positioning: The proposed work differs by combining extensible representations with global inter-class dependencies across sessions and non-training prototype updates.This contrast is stated relative to topology-preservation approaches for FSCIL.
- Few-Shot Learning: Few-shot learning methods include metric-based, meta-based, and augmentation-based approaches for adapting to novel classes.These categories respectively learn similarity functions, adaptation strategies, or synthetic novel-class data.
- Few-Shot Learning: Some few-shot methods model attention or graph-based dependencies to generate new-class weights while retaining base-class accuracy.These methods target rapid adaptation without forgetting initial classes.
3. Problem Description
FSCIL trains across disjoint class-incremental sessions, with abundant data initially and only a few samples per class thereafter. Evaluation tests recognition across all classes seen by each session.
- Task Definition: FSCIL uses a continuous stream of training sets X1, X2, · · · Xn with disjoint label sets across incremental sessions.Each session introduces classes not present in the other sessions.
- Task Definition: The first session has sufficient samples, whereas subsequent sessions provide only a few samples, such as five per class.This sampling regime defines the few-shot incremental setting.
- Evaluation: At session i, classification accuracy is measured on test set Zi containing classes from all label sets seen through that session.The evaluation therefore spans both previously learned and newly introduced classes.
4. Method
The method combines randomly generated incremental episodes with prototype refinement to learn extensible representations and preserve relations between old and new classes.
- Standard Learning Paradigm: The training objective jointly optimizes feature representations, prototypes, and metric parameters under classification supervision.The parameter set includes the feature extractor, learnable prototypes, and metric parameters; classification loss usually denotes cross-entropy.
- Random Episode Selection: Random episode selection generates N-way K-shot incremental episodes to force features to adapt to varied simulated incremental processes.The simulated process targets recognition of all seen classes rather than only the newly sampled classes.
- Dynamic Relation Projection: Dynamic relation projection transforms new-class embeddings and old prototypes into a shared latent space before computing their relations.The resulting relation matrix is used as a transition coefficient for prototype refinement.
- Dynamic Relation Projection: Prototype refinement uses relations between old classes and new classes to maintain dependencies while enhancing discrimination of new classes.The method dynamically moves prototypes toward preserving existing knowledge and improving new-class discrimination.
- Optimization: Unlike standard learning, the representation is encouraged to support classification of existing classes and beneficial updates for subsequent incremental tasks.Prototypes are jointly optimized with selected representations while satisfying mutual relation projection.
5. Experiment
Experiments evaluate the method on three benchmark datasets, analyze its components and training conditions, and compare it with state-of-the-art and classical baselines. Results show gains from extensible representations and prototype refinement, with improved accuracy across datasets and incremental sessions.
- Dataset and Settings: Experiments use CIFAR-100, MiniImageNet, and CUB200 under shared class-incremental settings with ResNet-18 and repeated sample testing.The evaluation reports session-wise and average accuracy, with five random incremental-sample selections averaged for each model.
- Ablation Study: About 5 percentage points of average overall improvement come from the extensible representation compared with standard representation.Without extensible representation, SPPR alone cannot play its role and performance drops by 4.24%.
- Ablation Study: SPPR improves performance by over 3 percent compared with fine-tuning, while adding fine-tuning to the full scheme does not boost performance.This comparison supports prototype updating through SPPR rather than an additional fine-tuning step.
- Analysis: Increasing prototype-update iterations reduces accuracy, whereas training remains broadly insensitive to the number of episodic ways and shots.Performance is slightly worse for 1-shot learning and drops slightly when the number of ways exceeds 5.
- Analysis: The 5-way-trained representation maintains nearly the same curve across test sessions, but reducing test images to 1 shot causes a significant final-result drop.Among tested prototype-update choices, the proposed method performs best.
- Comparison with SOTA: The method surpasses the state-of-the-art by over 13 percentage points on CIFAR-100, 17 percentage points on MiniImageNet, and over 11 percentage points on CUB200.On CIFAR-100 and MiniImageNet, accuracy is higher at all sessions and attenuates more slowly; confusion matrices also indicate stronger representation and prototypes across classes.
6. Conclusion
The paper proposes an incremental prototype learning scheme for FSCIL that combines extensible feature representation learning with prototype reorganization, achieving superior performance and adaptability to state-of-the-art methods.
- The method combines random episode selection to enhance feature representation extensibility with self-promoted prototype refinement to reorganize prototypes.
- The scheme incorporates incremental classes with few samples into recognition while preserving adaptability across incremental tasks.
- The model outperforms state-of-the-art methods in both performance and adaptability.