Source-linked AI summary
Few-Shot Class-Incremental Learning
Xiaoyu Tao, Xiaopeng Hong, Xinyuan Chang, Songlin Dong, Xing Wei, Yihong Gong
TL;DR
FSCIL asks CNNs to learn new classes from very few labelled samples without forgetting old classes. TOPIC uses a neural gas network to preserve and adapt feature-space topology, and it substantially outperforms state-of-the-art CIL methods across three datasets.
Problem
FSCIL requires incremental learning from very few new-class samples while preserving recognition of previously learned classes.
Method
TOPIC represents CNN feature-space knowledge with neural gas, stabilizing its topology for old classes and adapting it to few-shot new classes.
Results
TOPIC substantially outperforms other state-of-the-art CIL methods on CIFAR100, miniImageNet, and CUB200.
Takeaways & Limitations
Topology-preserving feature-space knowledge representation supports FSCIL with negligibly small memory overhead.
Takeaways & Limitations
Some CIL baseline components are less effective for FSCIL because cosine distance can discard feature-magnitude patterns and bias correction requires many training samples.
Abstract
from arXiv · showhide
The ability to incrementally learn new classes is crucial to the development of real-world artificial intelligence systems. In this paper, we focus on a challenging but practical few-shot class-incremental learning (FSCIL) problem. FSCIL requires CNN models to incrementally learn new classes from very few labelled samples, without forgetting the previously learned ones. To address this problem, we represent the knowledge using a neural gas (NG) network, which can learn and preserve the topology of the feature manifold formed by different classes. On this basis, we propose the TOpology-Preserving knowledge InCrementer (TOPIC) framework. TOPIC mitigates the forgetting of the old classes by stabilizing NG's topology and improves the representation learning for few-shot new classes by growing and adapting NG to new training samples. Comprehensive experimental results demonstrate that our proposed method significantly outperforms other state-of-the-art class-incremental learning methods on CIFAR100, miniImageNet, and CUB200 datasets.
1. Introduction
FSCIL addresses the practical need to learn emerging classes from very few examples while retaining old-class knowledge. TOPIC represents feature-space topology with neural gas and uses topology stabilization and adaptation to mitigate forgetting and improve few-shot learning.
- Motivation: FSCIL requires CNNs to learn new classes from very few labelled examples while retaining previously learned classes.The setting is motivated by dynamic environments where users provide limited annotations for new classes.
- Challenge: Limited-data finetuning causes both catastrophic forgetting of old classes and overfitting to new classes.These effects respectively degrade old-class performance and generalization to test samples.
- TOPIC: TOPIC uses a neural gas network to represent feature-space topology rather than constraining only CNN output logits.The framework stabilizes old topology with anchor loss and adapts the feature space for few-shot new classes.
- TOPIC: TOPIC grows neural gas to adapt to changes in feature space and optimizes preservation of old knowledge alongside representation learning for new classes.Its two objectives address forgetting and overfitting to few-shot new classes.
2. Related Work
Related work addresses incremental learning through output distillation, rehearsal, architectural changes, and regularization, but FSCIL differs by limiting new-class data and withholding old training data. TOPIC instead constrains the CNN feature space with neural gas.
- Class-Incremental Learning: CIL typically uses knowledge distillation and external memory to preserve old-class information while learning a unified classifier.Examples include exemplar-based iCaRL and end-to-end EEIL with a distillation loss.
- Class-Incremental Learning: TOPIC targets the more difficult FSCIL setting, where new classes have limited training samples, by constraining feature space rather than network outputs.This differs from conventional CIL approaches that constrain output logits.
- Multi-Task Incremental Learning: Multi-task incremental methods use rehearsal, architectural, or regularization strategies for sequences of disjoint tasks.Rehearsal may store exemplars or generate old-task samples, while architectural methods manipulate network structure.
- Multi-Task Incremental Learning: Multi-task methods are excluded from FSCIL comparisons because FSCIL uses a single-task multi-class scenario without task identities.The paper treats FSCIL as a practical multi-class CIL case.
- Dynamic Few-Shot Learning: Few-shot learning methods often rely on base-class data to sample meta-learning tasks, which is unavailable during FSCIL incremental stages.Consequently, these methods cannot be directly applied to FSCIL.
3. Few-Shot Class-Incremental Learning
This section defines FSCIL and presents TOPIC’s neural-gas representation for learning new classes while preserving previously learned feature-space structure. Its stabilization and adaptation mechanisms address forgetting and overfitting through topology preservation, anchor loss, and min-max loss.
- Problem: FSCIL incrementally learns disjoint new classes from few-shot training sets while testing one unified classifier on all encountered classes.Only the current session’s training set is available during each incremental session.
- Baseline limitations: Knowledge distillation can suffer from old/new class imbalance and an unfavorable trade-off between learning new classes and preserving old-class outputs.The paper therefore abandons distillation and instead constrains CNN feature space, which it argues contains richer information than output logits.
- Neural-gas representation: TOPIC represents CNN’s feature space with a neural gas graph whose centroid nodes partition the space into regions and whose edges encode neighborhood relations.Competitive Hebbian learning adapts centroids to inputs, while node labels and assigned samples support supervised incremental learning.
- Incremental learning: During incremental sessions, TOPIC grows the graph with new nodes and edges, stabilizes old nodes, and adapts new nodes to improve discrimination without overfitting few-shot data.For each new class, it inserts k < K nodes and updates their centroids and edges.
- Less-forgetting stabilization: Anchor loss penalizes shifts of old NG centroids using inverse variance, relaxing constraints on high-variance dimensions to preserve useful shared information.The anchor-loss effect is to avoid severe distortion of feature-space topology.
- Less-overfitting adaptation: Min-max loss pulls each new-class feature toward its matching centroid while pushing that centroid away from differently labeled neighboring centroids by a margin.This constrains local feature geometry so samples are less likely to activate neighbors belonging to other classes.
4. Experiment
TOPIC is evaluated against adapted CIL baselines across three datasets, CNN backbones, and few-shot incremental settings. It consistently outperforms competing methods, with gains attributed to preserving old-class knowledge and improving new-class representation learning.
- Experimental settings: Experiments cover CIFAR100 and miniImageNet with 5-way 5-shot sessions, and CUB200 with 10-way 5-shot sessions.CIFAR100 and miniImageNet use 60 base and 40 new classes; CUB200 uses 100 base and 100 new classes split across ten sessions.
- Baselines: TOPIC is compared with iCARL, EEIL, and NCM, while BiC is excluded because its bias-correction model requires impractically many validation samples.The evaluation uses QuickNet and ResNet18 baselines, with ResNet18 evaluated on all three datasets.
- Comparative results: Across three datasets, QuickNet and ResNet18, TOPIC outperforms state-of-the-art methods at every encountered session and remains closest to the Joint-CNN upper bound.Its superiority becomes more significant as the incremental sequence grows.
- Comparative results: 13.52%: TOPIC’s maximum CIFAR100 improvement over the two state-of-the-art methods, reaching 24.17% with QuickNet and 29.37% with ResNet18.The second-best accuracies are 19.50% and 15.85%, respectively.
- Comparative results: 4.84%: TOPIC’s maximum miniImageNet improvement over EEIL*, with final accuracies of 18.36% using QuickNet and 24.42% using ResNet18.EEIL* reaches 13.59% and 19.58%, respectively.
- Comparative results: 26.28%: TOPIC’s final CUB200 accuracy with ResNet18, exceeding EEIL* at 22.11% by up to 4.17%.The comparison uses the 10-way 5-shot setting after tuning competing methods’ learning rates.
- Ablation study: AL improves over its simplified version by up to 1.49%, while AL-MML exceeds AL by up to 5.85%; distillation instead degrades performance.The results support combining forgetting mitigation with representation learning for few-shot new classes.
- Ablation study: NG nodes consistently outperform randomly selected exemplars under different memory sizes in CIFAR100 experiments with ResNet18.The NG representation is used to preserve knowledge in the CNN feature space.
5. Conclusion
TOPIC preserves CNN feature-space knowledge for FSCIL by maintaining the topology of class feature manifolds with a neural gas network. It mitigates forgetting and improves few-shot new-class representation learning, outperforming state-of-the-art CIL methods across three datasets with negligible memory overhead.
- Framework: TOPIC uses a neural gas network to maintain the topological structure of the feature manifold formed by different classes.The framework represents knowledge in CNN feature space.
- Framework: TOPIC mitigates forgetting of old classes and improves representation learning for few-shot new classes.It stabilizes the neural gas topology while growing and adapting it to new training samples.
- Results: TOPIC substantially outperforms state-of-the-art CIL methods on CIFAR100, miniImageNet, and CUB200.The experiments cover all three named datasets.
- Results: TOPIC incurs a negligibly small memory overhead.