[Paper Review] Novel Class Discovery: an Introduction and Key Concepts

Colin Troisemaine, Vincent Lemaire+4 more|arXiv (Cornell University)|Feb 22, 2023

Machine Learning and Algorithms4 citations

TL;DR

This paper provides a comprehensive survey of Novel Class Discovery (NCD), a weakly supervised learning setting where models learn from labeled known classes and unlabeled unknown classes during training. It introduces a taxonomy of NCD methods based on knowledge transfer—two-stage versus one-stage learning—and highlights key techniques like pseudo-labeling, self-supervised learning, and contrastive learning, while distinguishing NCD from related fields like zero-shot learning and open-world learning.

ABSTRACT

Novel Class Discovery (NCD) is a growing field where we are given during training a labeled set of known classes and an unlabeled set of different classes that must be discovered. In recent years, many methods have been proposed to address this problem, and the field has begun to mature. In this paper, we provide a comprehensive survey of the state-of-the-art NCD methods. We start by formally defining the NCD problem and introducing important notions. We then give an overview of the different families of approaches, organized by the way they transfer knowledge from the labeled set to the unlabeled set. We find that they either learn in two stages, by first extracting knowledge from the labeled data only and then applying it to the unlabeled data, or in one stage by conjointly learning on both sets. For each family, we describe their general principle and detail a few representative methods. Then, we briefly introduce some new related tasks inspired by the increasing number of NCD works. We also present some common tools and techniques used in NCD, such as pseudo labeling, self-supervised learning and contrastive learning. Finally, to help readers unfamiliar with the NCD problem differentiate it from other closely related domains, we summarize some of the closest areas of research and discuss their main differences.

Motivation & Objective

• To formally define the Novel Class Discovery (NCD) problem and clarify its setup, distinguishing it from related domains like zero-shot learning and open-world learning.
• To categorize existing NCD methods into two-stage and one-stage frameworks based on how knowledge is transferred from labeled known classes to unlabeled unknown data.
• To identify and explain core techniques commonly used in NCD, such as pseudo-labeling, self-supervised learning, and contrastive learning.
• To clarify the distinctions between NCD and closely related tasks such as open-set recognition, out-of-distribution detection, and generalized category discovery.
• To highlight open challenges, including the need for unified benchmarks, automatic cluster number estimation, and adaptation of NCD methods to non-image data types.

Proposed method

• Proposes a taxonomy of NCD methods based on knowledge transfer strategy: two-stage (first learn from known classes, then apply to unknowns) and one-stage (joint learning on both sets).
• Reviews representative methods in each family, emphasizing how they leverage labeled data to improve clustering of unlabeled unknown classes.
• Introduces and explains key techniques such as pseudo-labeling, where model predictions on unlabeled data are used as supervision, and contrastive learning to improve feature representation.
• Applies self-supervised learning pretraining to enhance feature quality, especially in image-based NCD, using data augmentation and invariance learning.
• Uses graph-based and spectral clustering principles to improve cluster quality when semantic coherence is critical.
• Proposes a two-stage pipeline as a baseline: first detect unknown samples using open-set recognition methods, then cluster them via NCD techniques.

Figure 1: The open-world scenario, where new classes appear during inference.

Experimental results

Research questions

• RQ1 How does Novel Class Discovery differ from related paradigms such as zero-shot learning, open-world learning, and out-of-distribution detection?
• RQ2 What are the key methodological families in NCD, and how do two-stage and one-stage approaches compare in terms of performance and risk of overfitting?
• RQ3 To what extent can self-supervised and contrastive learning techniques improve feature representations for unknown class discovery?
• RQ4 What are the main challenges in adapting NCD methods to non-image data, such as tabular or text data, and how might current image-specific techniques be generalized?
• RQ5 How can the number of unknown classes be estimated automatically, and what impact does this assumption have on NCD performance?

Key findings

• One-stage NCD methods have largely replaced two-stage approaches due to reduced risk of overfitting on known classes, despite the latter's initial popularity.
• Unsupervised clustering alone fails on complex datasets due to assumptions about cluster shape, similarity metrics, and data distribution, often grouping by spurious features like image color rather than semantic content.
• Self-supervised and contrastive learning techniques significantly improve feature representations in image-based NCD, enabling better clustering of unknown classes.
• Pseudo-labeling is a critical component in one-stage NCD frameworks, and its accuracy strongly influences overall performance, indicating room for improvement through better labeling strategies.
• The field remains predominantly focused on image data due to the reliance on image-specific techniques like data augmentation and CNN-based architectures, with limited extension to other modalities.
• A unified benchmark and standardized evaluation protocol are urgently needed, as dataset splits and class splits significantly affect NCD problem difficulty and model performance comparisons.

Figure 2: The Novel Class Discovery scenario, where both labeled data of known classes and unlabeled data of unknown classes are available during training.