Divide and Conquer: Static-Dynamic Collaboration for Few-Shot Class-Incremental Learning

TL;DR

The paper tackles FSCIL by addressing the stability-plasticity trade-off through Static-Dynamic Collaboration (SDC), which splits learning into a Static Retaining Stage that locks in base knowledge via a static projector and a Dynamic Learning Stage that adds a lightweight dynamic projector to adapt to new classes. The two projectors are fused to classify both old and new concepts, aided by a growing prototype memory that preserves representative features. An information-theoretic analysis underpins the design, and extensive experiments on four benchmarks, including a real-world dataset, demonstrate state-of-the-art performance and robustness across incremental sessions. The approach remains simple, scalable, and effective, offering a practical pathway to continual learning with limited data per task.

Abstract

Few-shot class-incremental learning (FSCIL) aims to continuously recognize novel classes under limited data, which suffers from the key stability-plasticity dilemma: balancing the retention of old knowledge with the acquisition of new knowledge. To address this issue, we divide the task into two different stages and propose a framework termed Static-Dynamic Collaboration (SDC) to achieve a better trade-off between stability and plasticity. Specifically, our method divides the normal pipeline of FSCIL into Static Retaining Stage (SRS) and Dynamic Learning Stage (DLS), which harnesses old static and incremental dynamic class information, respectively. During SRS, we train an initial model with sufficient data in the base session and preserve the key part as static memory to retain fundamental old knowledge. During DLS, we introduce an extra dynamic projector jointly trained with the previous static memory. By employing both stages, our method achieves improved retention of old knowledge while continuously adapting to new classes. Extensive experiments on three public benchmarks and a real-world application dataset demonstrate that our method achieves state-of-the-art performance against other competitors.

Figures (3)

• Figure 1: (a) The setting of few-shot class-incremental learning, which aims to recognize new classes without forgetting old ones. (b) The phenomenon of imbalance between preserving old classes and adapting to new classes.
• Figure 2: Illustration of our framework (SDC). Left: The training process in the static retaining stage (session $0$). We train a model on the dataset $D^{(0)}$, which contains a backbone $\phi_b$ and a classifier $\phi_c^{(0)}$. The classifier consists of a projector $\delta_B$ and a fully connected layer $\delta_f$ for classification. Right: The training process in the dynamic learning stage (session $t (t>0)$). Before training, we freeze the backbone, retain the previous projector as a static projector $\delta_B$, and train a new dynamic projector $\delta_I^{(t)}$ with the assistance of a memory $\textbf{M}^{(t)}$ and gain the classifier $\phi_c^{(t)}$.
• Figure 3: Representation visualization with t-SNE 2008Visualizing, which uses session $0$ and session $1$ on CIFAR100 as an example. We randomly select 25 examples over 5 base classes and 1 incremental class to show the model effect. Symbols '$\bullet$' and '$\blacktriangle$' represent examples of base classes and incremental classes, respectively. (a) visual features of base classes from session $0$. (b) , (c), and (d) are visual features of all classes when $\alpha = 0$, $\alpha = 0.5$, and $\alpha = 1$ from session $1$, respectively.