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

Zhong Ji, Rongshuai Wei, Jingren Liu, Yanwei Pang, Jungong Han

---

🔥 News

• [2026-03-09] Repository initialized.
• [2026-03-09] README released.
• [2026-03-13] Training and evaluation code will be released.
• [2026-03-13] Pretrained checkpoints will be released.
• [2026-03-13] Reproducibility guide will be released.

---

🧩 Overview

This repository provides the official implementation for:

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

Self-Explainable Models (SEMs) rely on Prototypical Concept Learning (PCL) to make visual recognition more interpretable, but they often struggle in few-shot settings due to limited supervision. To address this issue, we propose a Few-Shot Prototypical Concept Classification (FSPCC) framework that systematically tackles two core challenges in low-data regimes:

• parametric imbalance between the backbone and the concept learning module,
• representation misalignment between feature embeddings and concept activations.

Our method introduces:

• a Mixture of LoRA Experts (MoLE) for parameter-efficient adaptation,
• cross-module concept guidance to align feature representations with prototypical concept activations,
• a multi-level feature preservation strategy to fuse spatial and semantic cues from multiple layers,
• a geometry-aware concept discrimination loss to reduce concept overlap and improve interpretability.

Experimental results on six popular benchmarks show that our method consistently outperforms existing self-explainable models by a clear margin in few-shot image classification.

---

✨ Highlights

• An interpretable few-shot image classification framework built on prototypical concept learning
• A parameter-efficient adaptation design using Mixture of LoRA Experts (MoLE)
• Cross-module concept guidance for better feature-concept alignment
• Multi-level feature preservation for stronger low-data representations
• Geometry-aware concept discrimination for clearer and less-overlapping concepts
• Strong results on CUB-200-2011, mini-ImageNet, CIFAR-FS, Stanford Cars, FGVC-Aircraft, and DTD

---

📁 Repository Structure

The current repository is organized as follows:

LE-ProtoPNet/
├── configs/                 # experiment configuration files
├── data/                    # dataset storage / preprocessing utilities
├── evaluation/              # evaluation metrics and testing utilities
├── models/                  # backbone networks, concept modules, MoLE implementation
├── shared_utils/            # common utilities (logging, helpers, training utils)
├── train_fsl.py             # few-shot training entry point
├── test_fsl.py              # few-shot evaluation script
└── env.yaml                 # conda environment specification

---

⚙️ Installation

1. Create Conda Environment

conda env create -f env.yaml
conda activate LEProtoPNet

---

📊 Supported Datasets

Our framework supports several commonly used few-shot image classification benchmarks:

• CUB-200-2011
• mini-ImageNet
• CIFAR-FS
• Stanford Cars
• FGVC Aircraft
• DTD (Describable Textures Dataset)

---

Dataset Directory Structure

After downloading the datasets, organize them as follows:

data/
├── cub/
│   ├── images
│   └── splits
├── mini_imagenet/
│   ├── images
│   └── splits
├── cifar_fs/
│   └── splits
├── stanford_cars/
├── fgvc_aircraft/
└── dtd/

Dataset split files should follow the standard few-shot meta-learning splits used in prior work.

---

🏋️ Training

The training pipeline is implemented in train_fsl.py.

python train_fsl.py

---

🧪 Evaluation

Evaluation is implemented in test_fsl.py.

python test_fsl.py

---

📜 Citation

If you find this repository useful for your research, please cite:

@article{ji2026interpretable,
  title={Interpretable Few-Shot Image Classification via Prototypical Concept-Guided Mixture of LoRA Experts},
  author={Ji, Zhong and Wei, Rongshuai and Liu, Jingren and Pang, Yanwei and Han, Jungong},
  journal={IEEE Transactions on Image Processing},
  year={2026},
  publisher={IEEE}
}

---

⭐ Star History

---

📬 Contact

For questions, suggestions, or collaborations:

• Email: jrl0219@tju.edu.cn
• GitHub Issues: Please open an issue in this repository.

---