Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CONCLAD: COntinuous Novel CLAss Detector

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

TL;DR

CONCLAD tackles continual novelty detection in settings with multiple novel classes and no oracle. It introduces an iterative multi-class uncertainty estimation framework that combines feature reconstruction error with per-class PCA transforms to produce robust novelty scores, enabling tiny supervision and selective pseudo-labeling for continual updates. Empirical results across Im21K-OOD, Eurosat, Plants, and Cifar100 demonstrate that CONCLAD consistently outperforms single-class detectors and semi-supervised baselines, especially as the number of novel classes per task increases. The approach reduces error propagation across task transitions and can adapt to multiple novel classes without collapsing them into a single new class, with code to be released.

Abstract

In the field of continual learning, relying on so-called oracles for novelty detection is commonplace albeit unrealistic. This paper introduces CONCLAD ("COntinuous Novel CLAss Detector"), a comprehensive solution to the under-explored problem of continual novel class detection in post-deployment data. At each new task, our approach employs an iterative uncertainty estimation algorithm to differentiate between known and novel class(es) samples, and to further discriminate between the different novel classes themselves. Samples predicted to be from a novel class with high-confidence are automatically pseudo-labeled and used to update our model. Simultaneously, a tiny supervision budget is used to iteratively query ambiguous novel class predictions, which are also used during update. Evaluation across multiple datasets, ablations and experimental settings demonstrate our method's effectiveness at separating novel and old class samples continuously. We will release our code upon acceptance.

CONCLAD: COntinuous Novel CLAss Detector

TL;DR

CONCLAD tackles continual novelty detection in settings with multiple novel classes and no oracle. It introduces an iterative multi-class uncertainty estimation framework that combines feature reconstruction error with per-class PCA transforms to produce robust novelty scores, enabling tiny supervision and selective pseudo-labeling for continual updates. Empirical results across Im21K-OOD, Eurosat, Plants, and Cifar100 demonstrate that CONCLAD consistently outperforms single-class detectors and semi-supervised baselines, especially as the number of novel classes per task increases. The approach reduces error propagation across task transitions and can adapt to multiple novel classes without collapsing them into a single new class, with code to be released.

Abstract

In the field of continual learning, relying on so-called oracles for novelty detection is commonplace albeit unrealistic. This paper introduces CONCLAD ("COntinuous Novel CLAss Detector"), a comprehensive solution to the under-explored problem of continual novel class detection in post-deployment data. At each new task, our approach employs an iterative uncertainty estimation algorithm to differentiate between known and novel class(es) samples, and to further discriminate between the different novel classes themselves. Samples predicted to be from a novel class with high-confidence are automatically pseudo-labeled and used to update our model. Simultaneously, a tiny supervision budget is used to iteratively query ambiguous novel class predictions, which are also used during update. Evaluation across multiple datasets, ablations and experimental settings demonstrate our method's effectiveness at separating novel and old class samples continuously. We will release our code upon acceptance.

Paper Structure

This paper contains 13 sections, 4 equations, 3 figures.

Table of Contents

  1. Introduction and Related Work
  2. Our Method
  3. Experiments
  4. Appendix
  5. Methodology Details
  6. Thresholds for Stopping the inner-loop
  7. How to define Ambiguity
  8. Measuring per-class uncertainty in CONCLAD's formulation
  9. Experimental Methodology Details
  10. Implementation Details for CONCLAD and Baselines
  11. Datasets:
  12. Baselines
  13. Additional Results

Figures (3)

  • Figure 1: (Left A.1,B.1) Continual Novelty Detection performance measured by AUROC at each task. The number of novel classes introduced per task is in parenthesis. Overall, CONCLAD (green) significantly over-performs baselines; (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 0.625% for CONCLAD, ER, PseudoER. Equivalent plots for Cifar100, Plants in appendix 4.3.
  • Figure 2: (Left) Continual Novelty Detection measured by AUROC; (Right) Ablations of CONCLAD. Supervision budget is 0.625% for Im21K, Eurosat and 1.25% for Plants, Cifar100
  • Figure 3: Results for Plants and Cifar100; (Left D.1,E.1) Continual Novelty Detection performance measured by AUROC at each task. The number of novel classes introduced per task is in parenthesis.(Center D.2,E.2) Results varying the supervision budget; (Right D.3,E.3) Results varying Novel Class Increment per task. For (left,right) Supervision budget is 1.25% for CONCLAD, ER, PseudoER. Overall, CONCLAD (green) significantly over-performs baselines