ConCM: Consistency-Driven Calibration and Matching for Few-Shot Class-Incremental Learning

March 3, 2026 · View on GitHub

Static Badge

PyTorch implementation of the ConCM Framework, ICLR 2026 🎉

Authors: Qinzhe Wang, Zixuan Chen, Keke Huang, Xiu Su, Chunhua Yang, Chang Xu

Abstract

Few-Shot Class Incremental Learning (FSCIL) is crucial for adapting to the complex open-world environments. Contemporary prospective learning-based space construction methods struggle to balance old and new knowledge, as prototype bias and rigid structures limit the expressive capacity of the embedding space. Different from these strategies, we rethink the optimization dilemma from the perspective of feature-structure dual consistency, and propose a Consistency-driven Calibration and Matching (ConCM) framework that systematically mitigates the knowledge conflict inherent in FSCIL. Specifically, inspired by hippocampal associative memory, we design a memory-aware prototype calibration that extracts generalized semantic attributes from base classes and reintegrates them into novel classes to enhance the conceptual center consistency of features. Further, to consolidate memory associations, we propose dynamic structure matching, which adaptively aligns the calibrated features to a session-specific optimal manifold space, ensuring cross-session structure consistency. This process requires no class-number priors and is theoretically guaranteed to achieve geometric optimality and maximum matching. On large-scale FSCIL benchmarks including mini-ImageNet, CIFAR100 and CUB200, ConCM achieves state-of-the-art performance, with harmonic accuracy gains of up to 3.41% in incremental sessions.

ConCM Framework

Results

Requirements

  • Python 3.10
  • PyTorch 2.5.1
  • tqdm
  • matplotlib
  • scikit-learn
  • numpy
  • pandas

Datasets and pretrained models

We follow FSCIL setting and use the same data index_list for training splits across incremental sessions. The datasets are made readily available by the authors of CEC in their github repository. Follow their provided instructions to download and unzip. Please make sure to overwrite the correct path in the shell script.

You can download the pretrained models here, following OrCo. Place the downloaded models under ./params and unzip it.

Extract prior knowledge

The ./prior folder contains the preprocessed <mini_imagenet/cub200>_part_prior_train.pickle files. You can use them directly in the train process without further processing.

If you want to rebuild the semantic prior knowledge of the dataset. Please download the file of glove_840b_300d and then perform:

   python ./prior/get_cub200_primitive_knowledge.py

   python ./prior/get_cifar100_primitive_knowledge.py

   python ./prior/get_miniimagenet_primitive_knowledge.py

Training

Find the scripts with the best hyperparameters under ./scripts.

As an example, to run the CUB200 experiment from the paper:

   chmod +x ./scripts/run_cub.sh
   ./scripts/run_cub.sh

To run the CIFAR100 experiment:

   chmod +x ./scripts/run_cifar100.sh
   ./scripts/run_cifar100.sh

To run the mini-ImageNet experiment:

   chmod +x ./scripts/run_miniimage.sh
   ./scripts/run_miniimage.sh

For the above experiments find the computed metrics available under: ./checkpoints/<dataset>/result.txt

Acknowledgment

Our project references the codes in the following repositories.