[Paper Review] Open-world machine learning: A review and new outlooks

Fei Zhu, Shijie Ma+5 more|arXiv (Cornell University)|Mar 4, 2024

Artificial Intelligence in Healthcare11 citations

TL;DR

A comprehensive survey of open-world machine learning (OWL) covering unknown rejection, novel class discovery, and class-incremental learning, with discussion of challenges, principles, and future directions.

ABSTRACT

Machine learning has achieved remarkable success in many applications. However, existing studies are largely based on the closed-world assumption, which assumes that the environment is stationary, and the model is fixed once deployed. In many real-world applications, this fundamental and rather naive assumption may not hold because an open environment is complex, dynamic, and full of unknowns. In such cases, rejecting unknowns, discovering novelties, and then continually learning them, could enable models to be safe and evolve continually as biological systems do. This article presents a holistic view of open-world machine learning by investigating unknown rejection, novelty discovery, and continual learning in a unified paradigm. The challenges, principles, and limitations of current methodologies are discussed in detail. Furthermore, widely used benchmarks, metrics, and performances are summarized. Finally, we discuss several potential directions for further progress in the field. By providing a comprehensive introduction to the emerging open-world machine learning paradigm, this article aims to help researchers build more powerful AI systems in their respective fields, and to promote the development of artificial general intelligence.

Motivation & Objective

• Introduce the open-world machine learning paradigm and its three core components: unknown rejection, novel class discovery, and class-incremental learning.
• Survey existing methods and categorize approaches for OWL under unknown rejection, NCD, and CIL.
• Identify challenges, limitations, and practical considerations to guide future research and applications.

Proposed method

• Synthesize and categorize recent advances in unknown rejection, including OOD detection and OSR, and discuss score-based, training-based, and outlier-aided approaches.
• Review novel class discovery techniques, including one-stage and multi-stage deep transfer clustering approaches, and discuss their transfer of knowledge from old to new classes.
• Summarize class-incremental learning in open-world settings, and discuss strategies to integrate discovery with incremental updates without catastrophic forgetting.

Experimental results

Research questions

• RQ1 What are the main tasks and life-cycle steps of open-world machine learning (unknown rejection, novel class discovery, and class-incremental learning)?
• RQ2 How do current methods address unknown rejection (OOD detection, OSR) and what are their strengths and limitations?
• RQ3 How can novel class discovery be effectively performed given labeled old classes and unlabeled unknowns?
• RQ4 What challenges arise when extending to class-incremental learning in open environments, and how can knowledge be accumulated without forgetting?

Key findings

• OWL integrates unknown rejection, novel class discovery, and continual learning to adapt to dynamic environments.
• OOD detection and OSR are central to unknown rejection, with multiple training and inference strategies (post-hoc, training-time, and outlier-based methods).
• Novel class discovery relies on transferring knowledge from labeled old classes to unlabeled novel classes, using both multi-stage and one-stage approaches.
• Generalized category discovery extends NCD by handling unlabeled data that includes both old and novel classes, often leveraging contrastive learning and non-parametric classifiers.
• Challenges include open-space risk, semantic similarity between old and new classes, and avoiding catastrophic forgetting during incremental updates.
• Evaluation relies on metrics such as AUROC for OOD detection and open-set recognition performance, with emphasis on reliable confidence estimation and robust representations.