ACIL: Active Class Incremental Learning for Image Classification

TL;DR

ACIL tackles the annotation-cost challenge in class incremental learning by introducing an active exemplar framework. Each episode offers a small labeled set and a large unlabeled pool, with a fixed exemplar budget $k$ carried forward to the next episode; exemplars are selected from $X^{U}_{n}$ and $E_{n-1}$ using a unified uncertainty-diversity objective and a budget split that respects class counts. The training loss combines a weighted cross-entropy on labeled data with a distillation term from the previous model, enabling retention of prior knowledge. Across six vision datasets, ACIL achieves competitive accuracy with state-of-the-art CIL methods while dramatically reducing annotation effort, validating its practicality for real-world continual learning scenarios; the approach remains robust to exemplar-budget variations and shows potential for extension to regression IL tasks.

Abstract

Continual learning (or class incremental learning) is a realistic learning scenario for computer vision systems, where deep neural networks are trained on episodic data, and the data from previous episodes are generally inaccessible to the model. Existing research in this domain has primarily focused on avoiding catastrophic forgetting, which occurs due to the continuously changing class distributions in each episode and the inaccessibility of the data from previous episodes. However, these methods assume that all the training samples in every episode are annotated; this not only incurs a huge annotation cost, but also results in a wastage of annotation effort, since most of the samples in a given episode will not be accessible to the model in subsequent episodes. Active learning algorithms identify the salient and informative samples from large amounts of unlabeled data and are instrumental in reducing the human annotation effort in inducing a deep neural network. In this paper, we propose ACIL, a novel active learning framework for class incremental learning settings. We exploit a criterion based on uncertainty and diversity to identify the exemplar samples that need to be annotated in each episode, and will be appended to the data in the next episode. Such a framework can drastically reduce annotation cost and can also avoid catastrophic forgetting. Our extensive empirical analyses on several vision datasets corroborate the promise and potential of our framework against relevant baselines.

Figures (3)

• Figure 1: Learning setup of ACIL. The numbers in parentheses denote the class labels present in the corresponding set. We consider the disjoint CIL setup 35_survey, where the samples is a given episode belong to different classes compared to the samples in all the previous episodes. Please refer to the text for more details.
• Figure 2: Performance analysis of ACIL. The AL baselines (Random, Coreset, BADGE) are shown with dotted lines; the CIL baselines (Finetuning, iCaRL, Rainbow, GDumb) and the proposed ACIL method are shown with solid lines. The error bars have been omitted from the Tiny ImageNet results for better visualization. Best viewed in color.
• Figure 3: Study of the exemplar set size (budget) on the SVHN dataset. The AL baselines (Random, Coreset, BADGE) are shown with dotted lines; the CIL baselines (Finetuning, iCaRL, Rainbow, GDumb) and the proposed ACIL method are shown with solid lines. Best viewed in color.