Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CUAL: Continual Uncertainty-aware Active Learner

Amanda Rios, Ibrahima Ndiour, Parual Datta, Jerry Sydir, Omesh Tickoo, Nilesh Ahuja

TL;DR

CUAL addresses the realistic problem of continual learning under evolving data where unlabeled post-deployment data may contain both old and novel classes, all under a tiny labeling budget. It introduces a continual active learner framework that couples a short-term AL loop with a long-term classifier, using per-class Feature Reconstruction Error (FRE) and class-specific PCA transforms to compute uncertainty and drive ambiguous-sample querying, including confident pseudo-labeling of novel classes and experience replay for long-term updates. The key contributions are the ambiguity-based querying strategy, integrated pseudo-labeling within continual AL, and extensive validation across diverse datasets and backbones showing superior performance under budget constraints. This approach reduces labeling costs while maintaining strong continual performance, enabling more practical open-world continual learning systems.

Abstract

AI deployed in many real-world use cases should be capable of adapting to novelties encountered after deployment. Here, we consider a challenging, under-explored and realistic continual adaptation problem: a deployed AI agent is continuously provided with unlabeled data that may contain not only unseen samples of known classes but also samples from novel (unknown) classes. In such a challenging setting, it has only a tiny labeling budget to query the most informative samples to help it continuously learn. We present a comprehensive solution to this complex problem with our model "CUAL" (Continual Uncertainty-aware Active Learner). CUAL leverages an uncertainty estimation algorithm to prioritize active labeling of ambiguous (uncertain) predicted novel class samples while also simultaneously pseudo-labeling the most certain predictions of each class. Evaluations across multiple datasets, ablations, settings and backbones (e.g. ViT foundation model) demonstrate our method's effectiveness. We will release our code upon acceptance.

CUAL: Continual Uncertainty-aware Active Learner

TL;DR

CUAL addresses the realistic problem of continual learning under evolving data where unlabeled post-deployment data may contain both old and novel classes, all under a tiny labeling budget. It introduces a continual active learner framework that couples a short-term AL loop with a long-term classifier, using per-class Feature Reconstruction Error (FRE) and class-specific PCA transforms to compute uncertainty and drive ambiguous-sample querying, including confident pseudo-labeling of novel classes and experience replay for long-term updates. The key contributions are the ambiguity-based querying strategy, integrated pseudo-labeling within continual AL, and extensive validation across diverse datasets and backbones showing superior performance under budget constraints. This approach reduces labeling costs while maintaining strong continual performance, enabling more practical open-world continual learning systems.

Abstract

AI deployed in many real-world use cases should be capable of adapting to novelties encountered after deployment. Here, we consider a challenging, under-explored and realistic continual adaptation problem: a deployed AI agent is continuously provided with unlabeled data that may contain not only unseen samples of known classes but also samples from novel (unknown) classes. In such a challenging setting, it has only a tiny labeling budget to query the most informative samples to help it continuously learn. We present a comprehensive solution to this complex problem with our model "CUAL" (Continual Uncertainty-aware Active Learner). CUAL leverages an uncertainty estimation algorithm to prioritize active labeling of ambiguous (uncertain) predicted novel class samples while also simultaneously pseudo-labeling the most certain predictions of each class. Evaluations across multiple datasets, ablations, settings and backbones (e.g. ViT foundation model) demonstrate our method's effectiveness. We will release our code upon acceptance.

Paper Structure

This paper contains 14 sections, 4 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction and Related Work
  2. Our Method
  3. Experiments
  4. Appendix
  5. CUAL Methodology Details
  6. Thresholds for Stopping the AL inner-loop and Pseudo-labeling
  7. How to define Ambiguity
  8. Measuring per-class uncertainty in CUAL $S^i(u)$ formulation
  9. Continual Model Update
  10. Experimental Methodology
  11. Implementation Details for CUAL and Baselines
  12. Datasets:
  13. Baselines
  14. Additional Results

Figures (3)

  • Figure 1: (Left A.1-C.1) Continual classification accuracy over continual tasks. The number of novel classes introduced per task is in parenthesis. CUAL over-performs other methods in this challenging setting. The Oracle (gray) is fully supervised ER.(Right A.2-C.2) Results varying the AL budget.
  • Figure 2: Continual Classification Results, averaged over all tasks. Left - Default CUAL and baselines ("ent" equates to Entropy). Right - CUAL ablations. Refer to main text for discussion. $^*$see apx 4.3.
  • Figure 3: Results for Cifar100; (Left D.1) Continual classification accuracy over continual tasks. The number of novel classes introduced per task is in parenthesis. Al budget is 1.25% as outlined in section 3.1. CUAL over-performs other methods in this challenging setting. The Oracle (gray) equates to classic fully supervised experience-replay.(Right D.2) Varying the Active supervision budget and it's effect on CUAL and AL baselines.