Learnable Distribution Calibration for Few-Shot Class-Incremental Learning.

Area of Science:

• Machine Learning
• Computer Vision

Background:

• Few-shot class-incremental learning (FSCIL) struggles with retaining knowledge of previous classes while learning new ones with minimal data.
• Existing methods often face challenges in accurately estimating distributions for both old and new classes.

Purpose of the Study:

• To propose a unified framework, Learnable Distribution Calibration (LDC), to address the dual challenges in FSCIL.
• To enhance the model's ability to memorize old class distributions and estimate new class distributions effectively.

Main Methods:

• Introduced a Parameterized Calibration Unit (PCU) that initializes class distributions using classifier vectors and a shared covariance matrix.
• Employed recurrent feature updates during base training to calibrate biased distributions.
• Utilized PCU during incremental learning to recover old class distributions and estimate/augment new class distributions.

Main Results:

• LDC demonstrated superior performance on CUB200, CIFAR100, and mini-ImageNet datasets, exceeding state-of-the-art by 4.64%, 1.98%, and 3.97% respectively.
• The method showed effectiveness in few-shot learning scenarios, validating its practical applicability.
• LDC requires no prior knowledge of class similarity, enhancing its flexibility.

Conclusions:

• LDC provides a robust and flexible solution for few-shot class-incremental learning.
• The proposed method effectively mitigates catastrophic forgetting and overfitting in incremental learning settings.
• LDC offers a unified framework for tackling key FSCIL challenges with fixed memory costs.