[Paper Review] Class Anchor Clustering: a Distance-based Loss for Training Open Set Classifiers

Dimity Miller, Niko Sünderhauf+2 more|arXiv (Cornell University)|Apr 6, 2020

Domain Adaptation and Few-Shot Learning4 citations

TL;DR

This paper proposes Class Anchor Clustering (CAC), a distance-based loss that trains deep neural networks to form tight, class-specific clusters around fixed, class-dependent anchor centers in the logit space. By explicitly optimizing for this clustering during training, CAC improves open set recognition performance, achieving a 2.4% AUROC gain on TinyImageNet without reducing accuracy, outperforming prior methods using learned centers.

ABSTRACT

In open set recognition, deep neural networks encounter object classes that were unknown during training. Existing open set classifiers distinguish between known and unknown classes by measuring distance in a network's logit space, assuming that known classes cluster closer to the training data than unknown classes. However, this approach is applied post-hoc to networks trained with cross-entropy loss, which does not guarantee this clustering behaviour. To overcome this limitation, we introduce the Class Anchor Clustering (CAC) loss. CAC is a distance-based loss that explicitly trains known classes to form tight clusters around anchored class-dependent cluster centres in the logit space. We show that training with CAC achieves state-of-the-art open set performance for distance-based open set classifiers on the standard benchmark datasets, with a 2.4% performance increase in AUROC on the challenging TinyImageNet, without sacrificing classification accuracy. We also show that our anchored class centres achieve higher open set performance than learnt class centres, particularly on object-based datasets and large numbers of training classes.

Motivation & Objective

• To address the limitation of existing open set classifiers that rely on post-hoc distance measurements in logit space without prior clustering optimization.
• To improve open set recognition performance by explicitly training known classes to form tight clusters around class-specific anchor centers during network training.
• To investigate whether fixed anchor centers outperform learned class centers in open set recognition, especially on object-based datasets and with large numbers of known classes.
• To maintain high zero-shot classification accuracy while significantly improving detection of unknown classes in open set scenarios.

Proposed method

• Introduces a novel distance-based loss function, Class Anchor Clustering (CAC), which minimizes the distance between each sample's logit representation and a fixed, class-dependent anchor center.
• Uses predefined, class-specific anchor centers—derived from known training data—instead of learning dynamic cluster centers during training.
• Applies the CAC loss in conjunction with standard cross-entropy loss to jointly optimize for both classification accuracy and intra-class compactness.
• Enforces that known class samples are pulled toward their respective anchor centers in the logit space, promoting tight clustering and improved separation from unknown classes.
• Employs a margin-based constraint to ensure that known class clusters remain distinct from the distribution of unknown class samples.
• Trains the network using a combined loss: cross-entropy for classification and CAC for cluster compactness, with anchor centers initialized from mean activations of training samples per class.

Experimental results

Research questions

• RQ1 Can explicitly training known classes to cluster around fixed anchor centers in logit space improve open set recognition performance compared to post-hoc distance-based methods?
• RQ2 Does using fixed class anchor centers outperform learning dynamic cluster centers in open set recognition, particularly on large-scale and object-based datasets?
• RQ3 To what extent does the CAC loss improve AUROC on standard benchmarks without degrading standard classification accuracy?
• RQ4 How does the performance of CAC vary with increasing numbers of known classes and on challenging datasets like TinyImageNet?

Key findings

• Training with the CAC loss achieves a 2.4% absolute improvement in AUROC on the TinyImageNet benchmark compared to prior state-of-the-art distance-based open set classifiers.
• The use of fixed class anchor centers leads to higher open set performance than learned cluster centers, especially on object-based datasets and with large numbers of training classes.
• CAC maintains high standard classification accuracy while significantly improving the model's ability to detect unknown classes.
• The improvement from CAC is most pronounced in challenging open set scenarios, demonstrating robustness across diverse dataset scales and class compositions.