Mamba-FSCIL: Dynamic Adaptation with Selective State Space Model for Few-Shot Class-Incremental Learning

Technical Summary of Mamba-FSCIL: Dynamic Adaptation with Selective State Space Model for Few-Shot Class Incremental Learning

Problem Statement
Few-Shot Class-Incremental Learning (FSCIL) aims to enable models to continuously learn novel classes from limited examples while retaining knowledge of previously learned base classes. Existing approaches face a critical dilemma regarding the stability-plasticity trade-off. Static adaptation methods employ fixed-capacity projectors shared across all sessions; however, updating these shared weights for new classes often overwrites parameters encoding base knowledge, leading to catastrophic forgetting. Conversely, dynamic adaptation methods with expanding architectures allocate new parameters for each session to preserve stability, but this results in unbounded model growth, increasing computational costs and limiting scalability. The core challenge is to achieve dynamic adaptation to new classes and preserve old knowledge without session-wise parameter expansion.

Methodology
The authors propose Mamba-FSCIL, a framework that leverages Selective State Space Models (SSMs), specifically Mamba, to enable input-dependent dynamic computation within a fixed-capacity architecture. Instead of expanding the model size, the method exploits Mamba's ability to generate core parameters ($B_i, C_i, \Delta_i$) on-the-fly based on the input content. The framework consists of three primary components:

1. Dual Selective SSM Projector:
   To structurally decouple stability and plasticity, the projector is divided into two parallel branches:

   • Base Selective SSM Branch ($g_{base}$): This branch acts as a "stability anchor." It is trained exclusively during the base session and then frozen. Unlike traditional static frozen modules, $g_{base}$ remains computationally dynamic; it continues to generate input-conditioned parameters even when frozen, providing stable, context-aware transformations that preserve foundational base-class representations.
   • Incremental Selective SSM Branch ($g_{inc}$): This branch acts as a "plasticity enabler." It is a single, fixed-capacity Mamba module reused across all incremental sessions. It leverages Mamba's input-dependent mechanism to capture residual feature shifts required for novel classes. By generating distinct dynamic parameters based on the input context, $g_{inc}$ adapts to new classes without adding new session-specific branches.

2. Class-Sensitive Selective Scan Mechanism:
   To ensure the dynamic adaptation of $g_{inc}$ is class-sensitive rather than merely content-aware, the authors introduce two regularization losses:

   • Suppression Loss ($\mathcal{L}_{supp}$): This loss minimizes the activation magnitude of the gating stream ($Z_{inc}$) for base-class inputs. By suppressing the incremental branch's contribution for known classes, it prevents interference with the frozen stability anchor. Conversely, it encourages a larger norm for novel-class inputs to facilitate adaptation.
   • Separation Loss ($\mathcal{L}_{sep}$): This loss minimizes the absolute cosine similarity between the averaged state-space parameters ($B, C, \Delta$) generated for base classes versus novel classes. This encourages the model to form distinct transformation pathways for new concepts, reducing representational overlap and enhancing plasticity.

3. Optimization Strategy:
   The framework utilizes a fixed Simplex Equiangular Tight Frame (ETF) classifier. During the base session, the backbone, identity branch, and $g_{base}$ are jointly optimized. In incremental sessions, the backbone and $g_{base}$ are frozen, and only $g_{inc}$ is optimized using a combination of Dot Regression (DR) loss, suppression loss, and separation loss. The system operates without storing or replaying raw images from previous sessions, relying instead on a compact feature-level memory of class-mean prototypes.

Key Contributions

• Mamba-FSCIL Framework: The first framework to apply Mamba's input-dependent dynamic computation to FSCIL, achieving class-sensitive adaptation within a fixed-capacity model.
• Dual Selective SSM Projector: A structural innovation that separates base and novel class processing via a frozen, computationally dynamic base branch and a reusable incremental branch, balancing stability and plasticity without parameter expansion.
• Class-Sensitive Selective Scan Mechanism: A novel regularization strategy using suppression and separation losses to explicitly guide the dynamic parameter generation of the incremental branch, ensuring distinct processing patterns for base and novel classes.
• State-of-the-Art Performance: Extensive experiments demonstrate that Mamba-FSCIL outperforms existing static and dynamic methods on miniImageNet, CIFAR-100, and CUB-200.

Experimental Results
The paper reports extensive evaluations on three benchmark datasets:

• miniImageNet: Using a ResNet-12 backbone, Mamba-FSCIL achieves an average accuracy (AVG) of 69.74% and a final session accuracy of 59.29%, surpassing the previous state-of-the-art (NC-FSCIL) by 1.92 and 0.98 points, respectively.
• CIFAR-100: With ResNet-12, the method achieves an AVG of 68.19% and a final accuracy of 57.52%, improving over NC-FSCIL and FeSSSS by significant margins.
• CUB-200: On this fine-grained dataset, Mamba-FSCIL achieves an AVG of 68.11% and a final accuracy of 61.52%. When using Transformer backbones (e.g., Swin-T, ViT-B/16), the method consistently achieves the highest or highly competitive performance, notably reducing the performance drop (PD) compared to expansion-based methods like DSN.
• Ablation Studies: Experiments confirm that the dual-branch design and class-sensitive losses are critical, with the full model outperforming variants that use static projectors, single SSM branches, or lack regularization. Qualitative analysis via activation maps and t-SNE visualizations further validates the model's ability to maintain stable base-class representations while dynamically adapting to novel classes.

Significance
The paper claims that Mamba-FSCIL offers a practical and scalable alternative to both parameter-sharing and parameter-expansion paradigms. By utilizing input-dependent dynamic computation, it resolves the stability-plasticity dilemma without the computational overhead of continually expanding model architectures. The work demonstrates that Selective State Space Models can effectively mitigate catastrophic forgetting in few-shot incremental settings by generating distinct, class-sensitive processing patterns for new data while preserving the integrity of established knowledge.