Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Uncertainty Quantification in Continual Open-World Learning

Amanda S. Rios, Ibrahima J. Ndiour, Parual Datta, Jaroslaw Sydir, Omesh Tickoo, Nilesh Ahuja

TL;DR

This work tackles continual open-world learning where unlabeled data streams contain both old and novel classes, a setting where traditional CL methods and fixed-oracle approaches fail. It introduces COUQ, an iterative uncertainty quantification framework built on per-class feature reconstruction error (FRE) with per-class PCA transforms, enabling robust novelty detection and enabling active labeling and pseudo-labeling in continual tasks. COUQ demonstrates strong performance across multiple datasets and backbones, maintaining reliable uncertainty estimates as new classes arrive and reducing labeling costs while mitigating error propagation. The approach enables practical deployment of open-world CL systems by replacing unrealistic oracle requirements with uncertainty-guided, data-efficient adaptation.

Abstract

AI deployed in the real-world should be capable of autonomously adapting to novelties encountered after deployment. Yet, in the field of continual learning, the reliance on novelty and labeling oracles is commonplace albeit unrealistic. This paper addresses a challenging and under-explored problem: a deployed AI agent that continuously encounters unlabeled data - which may include both unseen samples of known classes and samples from novel (unknown) classes - and must adapt to it continuously. To tackle this challenge, we propose our method COUQ "Continual Open-world Uncertainty Quantification", an iterative uncertainty estimation algorithm tailored for learning in generalized continual open-world multi-class settings. We rigorously apply and evaluate COUQ on key sub-tasks in the Continual Open-World: continual novelty detection, uncertainty guided active learning, and uncertainty guided pseudo-labeling for semi-supervised CL. We demonstrate the effectiveness of our method across multiple datasets, ablations, backbones and performance superior to state-of-the-art.

Uncertainty Quantification in Continual Open-World Learning

TL;DR

This work tackles continual open-world learning where unlabeled data streams contain both old and novel classes, a setting where traditional CL methods and fixed-oracle approaches fail. It introduces COUQ, an iterative uncertainty quantification framework built on per-class feature reconstruction error (FRE) with per-class PCA transforms, enabling robust novelty detection and enabling active labeling and pseudo-labeling in continual tasks. COUQ demonstrates strong performance across multiple datasets and backbones, maintaining reliable uncertainty estimates as new classes arrive and reducing labeling costs while mitigating error propagation. The approach enables practical deployment of open-world CL systems by replacing unrealistic oracle requirements with uncertainty-guided, data-efficient adaptation.

Abstract

AI deployed in the real-world should be capable of autonomously adapting to novelties encountered after deployment. Yet, in the field of continual learning, the reliance on novelty and labeling oracles is commonplace albeit unrealistic. This paper addresses a challenging and under-explored problem: a deployed AI agent that continuously encounters unlabeled data - which may include both unseen samples of known classes and samples from novel (unknown) classes - and must adapt to it continuously. To tackle this challenge, we propose our method COUQ "Continual Open-world Uncertainty Quantification", an iterative uncertainty estimation algorithm tailored for learning in generalized continual open-world multi-class settings. We rigorously apply and evaluate COUQ on key sub-tasks in the Continual Open-World: continual novelty detection, uncertainty guided active learning, and uncertainty guided pseudo-labeling for semi-supervised CL. We demonstrate the effectiveness of our method across multiple datasets, ablations, backbones and performance superior to state-of-the-art.

Paper Structure

This paper contains 29 sections, 4 equations, 3 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Contribution:
  3. Problem Statement - Background
  4. 2.2. Other Relevant Solutions:
  5. 2.2.1. Learning from Novelties:
  6. 2.2.2. Continual Learning:
  7. 2.2.3. Active Learning:
  8. Our Approach: COUQ
  9. Choice of elemental uncertainty metric:
  10. Algorithmic Steps
  11. Initial training and deployment:
  12. Continual Learning and Adaption:
  13. $i=0$: Initializing COUQ.
  14. $i>0$: Iterative Training.
  15. Application to Continual Novelty Detection
  16. ...and 14 more sections

Figures (3)

  • Figure 1: A general open-world, continual learning pipeline. Uncertainty estimation plays a crucial role in various points.
  • Figure 2: (Left A.1,B.1) AUROC of Novelty Detection at each continual task. Number of novel classes per task is in parenthesis. COUQ (green) clearly outperforms baselines both in both semi-supervised (solid line) and unsupervised versions (dashedline); (Center A.2,B.2) Results varying the supervision budget; (Right A.3,B.3) Results varying Novel Class Increment per task. For (left,right) Supervision budget is 1.25% and all plots show results implemented with a Resnet50 backbone. Equivalent plots for other datasets in appendix.
  • Figure 3: (Row 1) Continual classification accuracy over continual tasks during Continual Open-World Learning. The number of novel classes introduced per task for each dataset is in parenthesis. (Row 2) Results varying AL budget.