Kernel PCA for Out-of-Distribution Detection

Authors: Kun Fang, Qinghua Tao, Kexin Lv, Mingzhen He, Xiaolin Huang, JIE YANG

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive theoretical and empirical results on multiple Oo D data sets and network structures verify the superiority of our KPCA detector in efficiency and efficacy with state-of-the-art detection performance. Extensive experiments verify the effectiveness of the devoted two kernels.
Researcher Affiliation Collaboration Kun Fang1 Qinghua Tao2 Kexin Lv3 Mingzhen He1 Xiaolin Huang1 Jie Yang1 1Shanghai Jiao Tong University 2ESAT-STADIUS, KU Leuven 3China Mobile Shanghai ICT Co., Ltd
Pseudocode Yes Algorithm 1 Kernel PCA for Out-of-Distribution Detection
Open Source Code Yes The source code of this work has been publicly released1. 1https://github.com/fanghenshaometeor/ood-kernel-pca
Open Datasets Yes Datasets Experiments are executed on the commonly-used small-scale CIFAR10 [39] and large-scale Image Net-1K benchmarks [40], following the settings in [7, 8].
Dataset Splits Yes Datasets Experiments are executed on the commonly-used small-scale CIFAR10 [39] and large-scale Image Net-1K benchmarks [40], following the settings in [7, 8]. Following the setups in KNN, for fair comparisons, we evaluate models trained via the standard cross-entropy loss and models trained via the supervised contrastive learning [49], and adopt the same checkpoints released by KNN: Res Net18 [50] on CIFAR10 and Res Net50 on Image Net-1K.
Hardware Specification Yes All the experiments are executed on 1 NVIDIA Ge Force RTX 3090 GPU.
Software Dependencies No The detection experiments are executed on the Image Net-1K benchmark with pre-trained Res Net50 and Mobile Net [51] checkpoints from Py Torch [52].
Experiment Setup Yes Following the setups in KNN, for fair comparisons, we evaluate models trained via the standard cross-entropy loss and models trained via the supervised contrastive learning [49], and adopt the same checkpoints released by KNN: Res Net18 [50] on CIFAR10 and Res Net50 on Image Net-1K. in the comparison experiments, we adopt M = 4m on CIFAR10 with m = 512 for Res Net18, and M = 2m on Image Net-1K with m = 2048 for Res Net50 and m = 1280 for Mobile Net. A common hyper-parameter in Co P and Co RP is the number of columns q of the dimensionality-reduction matrix U Φ q . Additional hyper-parameters of Co RP include the bandwidth γ of the Gaussian kernel and the number of RFFs M.