CI-CBM: Class-Incremental Concept Bottleneck Model for Interpretable Continual Learning

1. Problem Statement

The paper addresses the challenge of Class-Incremental Learning (CIL), specifically in the Exemplar-Free (EFCIL) setting.

• The Core Issue: Deep learning models suffer from catastrophic forgetting, where learning new classes causes the model to forget previously learned knowledge.
• The Constraint: In the EFCIL setting, the model cannot store any raw data (exemplars) from previous phases due to privacy or storage limitations. It must learn new classes sequentially without access to old data.
• The Gap: Existing solutions for EFCIL often rely on "black-box" architectures that sacrifice interpretability. Conversely, existing interpretable models (like Concept Bottleneck Models or CBMs) struggle in CIL settings, often suffering from significant accuracy drops compared to black-box models or failing to maintain concept consistency across phases.

2. Methodology: CI-CBM

The authors propose CI-CBM (Class-Incremental Concept Bottleneck Model), a framework that extends the Label-free Concept Bottleneck Model (LF-CBM) to handle incremental learning without exemplars. The architecture consists of a frozen backbone, a concept bottleneck layer, and a sparse prediction layer.

The method operates through three primary modules across incremental phases:

A. Incremental Concept Set Expansion (Module I)

• Concept Generation: For new classes introduced in a phase, concepts are automatically generated using Large Language Models (LLMs, e.g., GPT-3) via specific prompts (e.g., "List features for recognizing X").
• Filtering: To prevent redundancy, the system uses text embeddings (Sentence-Transformer/CLIP) to filter out concepts that are too similar to existing class names or duplicate existing concepts.
• Result: A growing, unique concept set $C_t$ that accumulates knowledge across phases.

B. Knowledge-Preserving Concept Learning (Module II)

• Challenge: As the concept set expands, fine-tuning the concept bottleneck weights ($W_C$) on new data risks "concept drift," where previously learned concepts lose their semantic meaning.
• Solution: The authors introduce Concept Regularization via Knowledge Distillation.
  • Before updating weights for new concepts, the model computes the activation scores of current concepts on the new phase data using the old weights ($W_{C}^{t-1}$).
  • During training of the new weights ($W_{C}^{t}$), a distillation loss is added to the alignment loss. This forces the new model's output for old concepts to remain close to the previous model's output, preventing drift while allowing the addition of new concept neurons.

C. Dynamic Adaptation & Pseudo-Concept Generation (Module III)

• Challenge: The prediction layer ($W_F$) must classify both new classes (with real data) and old classes (with no data).
• Solution: The authors generate Pseudo-Features and Pseudo-Concepts for old classes.
  1. Pseudo-Feature Generation: For each old class $c_p$, the system identifies the nearest new class $c_n$ based on centroid similarity in the backbone feature space. It then shifts the feature distribution of $c_n$ to match the mean of $c_p$:
     $$\hat{f}(c_p) = f(c_n) - \mu(c_n) + \mu(c_p)$$
  2. Pseudo-Concept Projection: These pseudo-features are projected through the current concept bottleneck ($W_C^t$) to generate Pseudo-Concepts.
  3. Sparse Training: The final sparse predictor ($W_F$) is trained using real concepts for new classes and pseudo-concepts for old classes. This allows the model to learn a unified classifier without storing any past data, only the class centroids.

3. Key Contributions

1. First Robust Interpretable EFCIL Model: CI-CBM is the first interpretable CIL model proven to be robust in both pretrained (e.g., ImageNet-pretrained ViT) and non-pretrained (trained from scratch) scenarios.
2. Performance Parity with Black-Boxes: Unlike previous interpretable methods that lag significantly in accuracy, CI-CBM achieves performance comparable to state-of-the-art black-box models, reducing the accuracy gap to negligible levels.
3. Novel Techniques:
   • Concept Regularization: Prevents semantic drift of learned concepts during incremental updates.
   • Pseudo-Concept Generation: A mechanism to simulate old class data in the concept space, enabling the training of the sparse classifier without exemplars.
4. Dual-Level Interpretability: The model provides:
   • Global Interpretability: Understanding which concepts drive decisions for specific classes (visualized via Sankey diagrams).
   • Local Interpretability: Explaining individual predictions by showing the contribution of specific concepts (e.g., "blue head" and "forked tail" for a Tree Swallow).

4. Experimental Results

The authors evaluated CI-CBM on seven datasets: CIFAR-10, CIFAR-100, CUB-200, TinyImageNet, Places365, ImageNet-Subset, and ImageNet.

• Accuracy Gains: CI-CBM outperforms previous interpretable CIL methods (ICICLE, IN2, CONCIL, CLG-CBM) by a significant margin, achieving an average accuracy gain of 36%.
  • On CIFAR-100, it improved by ~39.5% over baselines.
  • On CUB, it improved by ~43% over baselines.
• Comparison to Black-Boxes:
  • In non-pretrained scenarios (ResNet-18/DeiT trained from scratch), CI-CBM achieves accuracy within 7.5% of the best unrestricted (black-box) SOTA methods.
  • In pretrained scenarios (ViT-Base/16), it is competitive with SOTA prompt-based methods, showing only a ~7% drop on CIFAR-100 compared to top black-box models.
• Ablation Studies:
  • Removing pseudo-concept generation causes accuracy to drop to near zero for old classes.
  • Removing concept regularization leads to concept drift and performance degradation.
  • Using SigLIP (instead of CLIP) for concept alignment yields slightly better accuracy and concept fidelity.
• Forgetting: The method demonstrates low catastrophic forgetting, maintaining high accuracy on old classes even after 20+ incremental phases.

5. Significance and Impact

• Bridging the Gap: CI-CBM successfully bridges the long-standing trade-off between interpretability and accuracy in continual learning. It proves that models can be transparent and human-understandable without sacrificing performance.
• Real-World Applicability: By operating in the Exemplar-Free setting, CI-CBM addresses critical privacy and storage constraints, making it suitable for real-world applications (e.g., healthcare, robotics) where data cannot be retained.
• Flexibility: The method's ability to function effectively whether the backbone is pretrained or trained from scratch makes it highly versatile for different deployment scenarios.
• Trustworthy AI: By maintaining consistent concept definitions across learning phases, CI-CBM ensures that the model's reasoning remains stable and auditable over time, a crucial requirement for high-stakes decision-making systems.

In conclusion, CI-CBM represents a major step forward in making continual learning systems both effective and explainable, offering a practical solution to the catastrophic forgetting problem without relying on data retention or sacrificing model transparency.