Interpretable Few-Shot Image Classification via Prototypical Concept-Guided Mixture of LoRA Experts

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Self-Explainable Models (SEMs) utilize Prototypical Concept Learning (PCL) for interpretable visual recognition.
• SEMs face performance degradation in data-scarce settings due to insufficient training data.

Purpose of the Study:

• To propose a Few-Shot Prototypical Concept Classification (FSPCC) framework to address limitations of SEMs in low-data regimes.
• To enhance both the accuracy and interpretability of visual recognition systems under few-shot conditions.

Main Methods:

• Leveraging a Mixture of LoRA Experts (MoLE) for parameter-efficient adaptation and balanced parameter allocation.
• Employing cross-module concept guidance for aligning feature representations with prototypical concept activation patterns.
• Incorporating multi-level feature preservation and a geometry-aware concept discrimination loss for disentangled decision boundaries.

Main Results:

• The FSPCC framework systematically mitigates parametric imbalance and representation misalignment in few-shot learning.
• Achieved relative accuracy gains of 4.2%-8.7% in 5-way 5-shot classification across six benchmarks.
• Demonstrated consistent outperformance compared to existing SEMs in low-data scenarios.

Conclusions:

• Coupling concept learning with few-shot adaptation effectively enhances both accuracy and interpretability in visual recognition.
• The proposed FSPCC framework offers a promising direction for developing more transparent and robust AI systems.
• The findings underscore the importance of addressing data scarcity for reliable and interpretable AI models.