Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
153 tokens/sec
GPT-4o
7 tokens/sec
Gemini 2.5 Pro Pro
45 tokens/sec
o3 Pro
4 tokens/sec
GPT-4.1 Pro
38 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Unseen Class Discovery in Open-world Classification (1801.05609v1)

Published 17 Jan 2018 in cs.LG and cs.AI

Abstract: This paper concerns open-world classification, where the classifier not only needs to classify test examples into seen classes that have appeared in training but also reject examples from unseen or novel classes that have not appeared in training. Specifically, this paper focuses on discovering the hidden unseen classes of the rejected examples. Clearly, without prior knowledge this is difficult. However, we do have the data from the seen training classes, which can tell us what kind of similarity/difference is expected for examples from the same class or from different classes. It is reasonable to assume that this knowledge can be transferred to the rejected examples and used to discover the hidden unseen classes in them. This paper aims to solve this problem. It first proposes a joint open classification model with a sub-model for classifying whether a pair of examples belongs to the same or different classes. This sub-model can serve as a distance function for clustering to discover the hidden classes of the rejected examples. Experimental results show that the proposed model is highly promising.

Citations (69)

Summary

  • The paper’s main contribution is a novel joint OCN-PCN model that integrates Gaussian-based outlier detection with hierarchical clustering to identify unseen classes.
  • It employs a one-vs-rest sigmoid layer to effectively reject unknown examples, surpassing existing methods like OpenMax in precision and recall on datasets such as MNIST.
  • The approach bridges supervised and unsupervised learning, enabling practical real-time class discovery for adaptive systems in dynamic open-world environments.

Unseen Class Discovery in Open-world Classification

The paper "Unseen Class Discovery in Open-world Classification" by Lei Shu, Hu Xu, and Bing Liu from the University of Illinois at Chicago addresses the challenge of extending the capabilities of open-world classification systems. These systems are required to classify known classes while rejecting and identifying examples from unseen classes that were not present during training—a task demanding both closed set recognition and novel class discovery.

The authors propose a robust method that leverages a joint model, integrating an Open Classification Network (OCN) with a Pairwise Classification Network (PCN), further enriched by hierarchical clustering techniques. The OCN is tasked with recognizing instances of seen classes, based on a unique one-vs-rest layer of sigmoid functions, diverging from standard softmax layers to enable rejection of unseen examples. The optimal rejection is achieved using a Gaussian-based outlier detection approach to tighten decision boundaries, thereby improving the system's ability to accurately leave out novel examples.

A major contribution of this research is the use of a PCN that employs a binary classification model to determine if examples come from the same class. The PCN learns from a multitude of pairs of examples drawn from known classes, subsequently serving as a distance function to discern intra-class similarities and inter-class differences. This ability to assess example similarity is pivotal for clustering tasks where the number of clusters (unseen classes) is initially unknown.

Experimental evaluations conducted on MNIST and EMNIST datasets evidence the effectiveness of the proposed framework. Notably, the framework outperforms existing methods such as OpenMax in rejecting examples from unseen classes, achieving significant precision and recall improvements. Furthermore, clustering experiments demonstrated that the hierarchical method guided by PCN successfully approximates the number of novel classes, showcasing a marked improvement over more traditional clustering methods like K-means, which requires pre-specification of clusters.

This paper advances the theoretical understanding of open-world learning by demonstrating the feasibility of discovering unseen classes through a transfer learning approach that bridges supervised and unsupervised learning paradigms. By elucidating a system capable of unsupervised class discovery post-rejection, the paper lays a foundation for adaptive learning systems in dynamic environments where continual class evolution is expected.

In future developments, improvements could focus on refining the threshold determination strategy for hierarchical clustering and enhancing transfer learning between seen and unseen classes to potentially accommodate a broader range of feature representations and application domains. Additionally, exploration into applying this methodology to other types of data beyond image classification could broaden its applicability and utility in open-world contexts.

This research offers significant implications for practical applications, particularly in areas where real-time adaptability to new and evolving classes is crucial. The proposed system's capability to automatically detect and learn new classes incrementally highlights its potential utility in surveillance systems, autonomous driving, and other AI systems that operate within open-world environments.

Youtube Logo Streamline Icon: https://streamlinehq.com