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

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

Domain Adaptation and Few-Shot Learning30 references4 citations

TL;DR

This paper proposes Class Anchor Clustering (CAC), a novel distance-based loss that trains deep neural networks to form tight, distinct clusters around fixed, class-specific centers in the logit space. By anchoring class centers instead of learning them during training, CAC improves open set recognition performance by 15.2% AUROC on TinyImageNet without sacrificing classification accuracy, especially excelling on object-based datasets and large-class scenarios.

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 centres in the logit space. We show that training with CAC achieves state-of-the-art performance for distance-based open set classifiers on all six standard benchmark datasets, with a 15.2% AUROC increase 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 measures in logit space after cross-entropy training, which does not guarantee tight clustering of known classes.
• To improve open set recognition by explicitly training known classes to cluster tightly around class-specific, fixed centers in the logit space.
• To investigate whether anchored class centers—fixed in advance—outperform learned class centers during training, especially on object-based datasets with high intra-class variation.
• To evaluate the scalability and robustness of anchored centers under increasing numbers of training classes.
• To develop a loss function that enhances open set performance without degrading standard classification accuracy.

Proposed method

• Introduce a new loss term, Class Anchor Clustering (CAC), which enforces known training samples to cluster tightly around predefined, class-specific centers in the logit space.
• Anchor class centers in the logit space using fixed, class-dependent values, eliminating the need to learn them during backpropagation.
• Combine CAC loss with standard cross-entropy loss for training stability, ensuring end-to-end optimization of both classification and clustering objectives.
• Use a distance-based objective that minimizes the L2 distance between each sample’s logit vector and its corresponding anchored class center.
• Apply the CAC loss during training to encourage spatial separation between known class clusters and to push unknown samples away from these clusters.
• Evaluate performance using AUROC on standard open set benchmarks, comparing anchored centers against learned centers under varying dataset and class complexity.

Experimental results

Research questions

• RQ1 Does training with anchored class centers in the logit space lead to better open set recognition performance than learning class centers during training?
• RQ2 How does the CAC loss improve the clustering of known classes in the logit space compared to cross-entropy training alone?
• RQ3 Does the use of anchored centers provide a performance advantage on object-based datasets with high intra-class variation, such as CIFAR10 and TinyImageNet?
• RQ4 How does the performance of CAC with anchored centers scale with increasing numbers of known classes?
• RQ5 Is the CAC loss robust to hyperparameter choices, particularly the anchor magnitude and loss weight?

Key findings

• Training with the CAC loss achieves state-of-the-art performance for distance-based open set recognition across all six standard benchmark datasets.
• On the challenging TinyImageNet dataset, CAC improves AUROC by 15.2% compared to prior methods, without any drop in classification accuracy.
• Anchored class centers outperform learned class centers by an average of 2.2% on object-based datasets (CIFAR10 variants and TinyImageNet) when using Center loss, and by 0.85% when using CAC loss.
• The performance gap between anchored and learned centers increases with the number of training classes, indicating that anchored centers scale better to large-class scenarios.
• Networks trained with anchored centers converge approximately 50% faster than those with learned centers, suggesting improved training stability and efficiency.
• The CAC loss is robust to hyperparameter variations, with both classification accuracy and AUROC varying by less than 4% across a wide range of anchor loss weights (0.05–0.8) and anchor magnitudes (5–20).