Boosting Open Set Recognition Performance through Modulated Representation Learning
Amit Kumar Kundu, Vaishnavi S Patil, Joseph Jaja
TL;DR
This paper introduces temperature-modulated representation learning with novel schedules, including a negative cosine schedule, to improve open set recognition by enabling richer feature exploration without extra computational cost.
Contribution
It proposes a set of temperature schedules for representation learning in OSR, enhancing model generalization and performance without additional overhead.
Findings
Improved OSR accuracy on challenging benchmarks
Enhanced closed set classification performance
Compatible with various existing loss functions
Abstract
The open set recognition (OSR) problem aims to identify test samples from novel semantic classes that are not part of the training classes, a task that is crucial in many practical scenarios. However, the existing OSR methods use a constant scaling factor (the temperature) to the logits before applying a loss function, which hinders the model from exploring both ends of the spectrum in representation learning -- from instance-level to semantic-level features. In this paper, we address this problem by enabling temperature-modulated representation learning using a set of proposed temperature schedules, including our novel negative cosine schedule. Our temperature schedules allow the model to form a coarse decision boundary at the beginning of training by focusing on fewer neighbors, and gradually prioritizes more neighbors to smooth out the rough edges. This gradual task switching leads…
Peer Reviews
Decision·ICLR 2026 Poster
The idea of temperature scheduling is novel and insightful. In particular, the negative cosine scheduling achieves a smooth transition of task switching by reverse-adjusting the temperature, which is more effective than fixed temperature or traditional cosine scheduling. The method does not require introducing additional regularization or generating synthetic samples, reducing computational costs and providing a concise yet efficient solution to the OSR problem.
Although comparisons with baseline methods (e.g., ARPL, SupCon) have been conducted, some recent OSR methods based on augmentation or multi-expert models (e.g., Wang et al. 2024) are not included. It is recommended to incorporate a broader range of comparisons in the rebuttal or future work to fully demonstrate the superiority. The performance improvement is limited on benchmarks with a small number of training classes (e.g., CIFAR), indicating that the method may be more suitable for scenarios
1. For open set recognition, Negative Cosine Schedule (NegCosSch) is proposed for temperature scheduling in softmax used in supervised contrastive learning. 2. The proposed variations of P-NegCosSch and M-NegCosSch together with exponential increase and linear increase generally outperforms the other 6 existing methods over 4 datasets. 3. The paper is generally well written.
1. The proposed method is a minor variation of the existing NegCosSch method for temperature scheduling. Hence, the novelty level is not high. 2. The P-NegCosSch and M-NegCosSch do not seem to perform better than the simpler exponential increase and linear increase. 3. The motivation for the periodic temperature schedule could be further discussed. Also, it does not seem to outperform the monotonic increasing version. 4. Part of the motivation is from semantic features vs instance/conte
+ This paper is well-written and easy to follow. I favor studies that derive methodological innovations from experimental observations. + The details provided in the Appendix are clear and crucial, enhancing the paper’s overall logical flow and strengthening the credibility of some experiments. + The empirical evaluation is extensive, covering multiple benchmarks and ablation studies.
+ The choice of $P$, $(\tau ^{+}, \tau ^{-})$ in the method is mainly heuristic. A more thorough theoretical analysis would improve the quality of the paper. + Some of the authors’ statements lack experimental or theoretical support. For example, the discussion regarding empty regions in the representation space, the significant computational overhead in existing methods, and the statement *'Therefore, the methods that demonstrate improvement on smaller datasets…'*(lines 096–098). + From my pe
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsFace and Expression Recognition · Anomaly Detection Techniques and Applications · Neural Networks and Applications