Few-Shot Class Incremental Learning via Robust Transformer Approach

TL;DR

This paper introduces ROBUSTA, a robust transformer-based method for few-shot class-incremental learning that effectively addresses overfitting and catastrophic forgetting, outperforming prior methods without data augmentation.

Contribution

The paper proposes a novel transformer-based framework with stochastic classifiers, delta parameters, and prototype rectification for improved few-shot class-incremental learning.

Findings

Outperforms prior methods on benchmark datasets

Effectively mitigates overfitting with stochastic classifiers and batch normalization

Handles catastrophic forgetting with delta parameters and prototype rectification

Abstract

Few-Shot Class-Incremental Learning presents an extension of the Class Incremental Learning problem where a model is faced with the problem of data scarcity while addressing the catastrophic forgetting problem. This problem remains an open problem because all recent works are built upon the convolutional neural networks performing sub-optimally compared to the transformer approaches. Our paper presents Robust Transformer Approach built upon the Compact Convolution Transformer. The issue of overfitting due to few samples is overcome with the notion of the stochastic classifier, where the classifier's weights are sampled from a distribution with mean and variance vectors, thus increasing the likelihood of correct classifications, and the batch-norm layer to stabilize the training process. The issue of CF is dealt with the idea of delta parameters, small task-specific trainable parameters while keeping the backbone networks frozen. A non-parametric approach is developed to infer the delta parameters for the model's predictions. The prototype rectification approach is applied to avoid biased prototype calculations due to the issue of data scarcity. The advantage of ROBUSTA is demonstrated through a series of experiments in the benchmark problems where it is capable of outperforming prior arts with big margins without any data augmentation protocols.