Unlocking Better Closed-Set Alignment Based on Neural Collapse for Open-Set Recognition

Authors: Chaohua Li, Enhao Zhang, Chuanxing Geng, Songcan Chen

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our theoretical analysis proves the validity of the proposed approach, and extensive experiments demonstrate that DEF achieves comparable or superior results with reduced computational resources on standard OSR benchmarks. Comprehensively experimental results indicate that DEF provides comparable or superior performance with less computational resources on standard OSR benchmarks.
Researcher Affiliation Academia 1College of Computer Science and Technology, Nanjing University of Aeronautics and Astronautics 2MIIT Key Laboratory of Pattern Analysis and Machine Intelligence 3Department of Computer Science, Hong Kong Baptist University EMAIL
Pseudocode Yes Pseudo-code are shown in Appendix F.
Open Source Code No The paper mentions code implementations for baselines in Appendix C but does not explicitly state that their own code (DEF) is released or available, nor does it provide a link to a repository for the methodology described.
Open Datasets Yes Following the protocol defined in (Neal et al. 2018) and the dataset splits specified in (Vaze et al. 2022; Li et al. 2024), we provide a summary of six standard benchmark datasets in OSR: SVHN,CIFAR10. SVHN(Netzer et al. 2011) and CIFAR10(Krizhevsky, Hinton et al. 2009) all contain 10 classes, with 6 classes randomly selected as closed set and the other 4 as open set. CIFAR+10,CIFAR+50. For these experiments, 4 classes from CIFAR-10 are selected as closed-set classes, while 10\50 classes from CIFAR-100(Krizhevsky, Hinton et al. 2009) are used as open-set classes. Tiny Image Net. Tiny Image Net is a subset derived from Image Net(Russakovsky et al. 2015) consisting of 200 classes. 20 known classes and the left 180 unknown classes are randomly sampled for evaluation. Datasets. The experiments involves two settings: 1) MNIST(Lake, Salakhutdinov, and Tenenbaum 2015) serves as the closed-set for training, while Omniglot (Lake, Salakhutdinov, and Tenenbaum 2015), MNIST-Noise and Noise are used as open-set during testing. 2) CIFAR10 is used as the closed-set dataset, with open-set data sampled from Image Net and LSUN(Yu et al. 2015).
Dataset Splits Yes Following the protocol defined in (Neal et al. 2018) and the dataset splits specified in (Vaze et al. 2022; Li et al. 2024), we provide a summary of six standard benchmark datasets in OSR: SVHN,CIFAR10. SVHN(Netzer et al. 2011) and CIFAR10(Krizhevsky, Hinton et al. 2009) all contain 10 classes, with 6 classes randomly selected as closed set and the other 4 as open set. CIFAR+10,CIFAR+50. For these experiments, 4 classes from CIFAR-10 are selected as closed-set classes, while 10\50 classes from CIFAR-100(Krizhevsky, Hinton et al. 2009) are used as open-set classes. Tiny Image Net. Tiny Image Net is a subset derived from Image Net(Russakovsky et al. 2015) consisting of 200 classes. 20 known classes and the left 180 unknown classes are randomly sampled for evaluation.
Hardware Specification No The paper does not provide any specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments.
Software Dependencies No The paper does not provide specific software dependency details, such as library names with version numbers (e.g., PyTorch 1.9, Python 3.8, CUDA 11.1).
Experiment Setup Yes In the training phase, we adopt a two-stage training strategy. Specifically, in the first stage, we train a feature extraction comprising an encoder network E ( ) and a projection network ψ( ). Both networks are optimized by supervised contrastive learning loss and our designed F-DEF loss on features of ψ(E (xi)). In the second stage, the learned E ( ) and ψ( ) are frozen and used to train a linear classifier f ( ) on the combination of cross entropy loss and our designed C-DEF loss. where α [0, 1] is sampled from the Beta distribution. where β [0, 1] is sampled from the Beta distribution. The runtime consumption (hours) of training DCTAU and our DEF for 600 epochs across four datasets. As shown in Figure 4, our method is relatively insensitive to the choice of α and β. (2)Threshold ϵ. We carried out the analysis on ϵ, and the results are provided in Appendix E.