RankFeat: Rank-1 Feature Removal for Out-of-distribution Detection

Authors: Yue Song, Nicu Sebe, Wei Wang

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive ablation studies and comprehensive theoretical analyses are presented to support the empirical results. Our Rank Feat establishes the state-of-the-art performance on the large-scale Image Net benchmark and a suite of widely used OOD datasets across different network depths and architectures.
Researcher Affiliation Academia Yue Song1, Nicu Sebe1, and Wei Wang2 1Department of Information Engineering and Computer Science, University of Trento, Italy 2Beijing Jiaotong University, China
Pseudocode No The paper describes the Power Iteration algorithm in text within Section 2, but it does not present it in a formal pseudocode block or algorithm environment.
Open Source Code Yes Code is publicly available via https://github.com/King James Song/Rank Feat.
Open Datasets Yes Datasets. In line with [26, 52, 27], we mainly evaluate our method on the large-scale Image Net-1k benchmark [6]. The large-scale dataset is more challenging than the traditional CIFAR benchmark [36] because the images are more realistic and diverse (i.e., 1.28M images of 1, 000 classes). For the OOD datasets, we select four testsets from subsets of i Naturalist [58], SUN [63], Places [70], and Textures [5].
Dataset Splits No The paper mentions using ImageNet-1k as the in-distribution (ID) training data for the model and other datasets as OOD test sets, but it does not specify explicit training/validation/test dataset splits (e.g., percentages or counts) for their experiments. It primarily evaluates OOD detection on pre-trained models.
Hardware Specification No The paper mentions 'Processing Time Per Image (ms)' and 'The test batch size is set as 16' but does not provide specific details about the hardware used, such as GPU/CPU models, memory, or cloud instance types.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., Python 3.x, PyTorch 1.x, CUDA 11.x).
Experiment Setup Yes Unless explicitly specified, we apply Rank Feat on the Block 4 feature by default. The main evaluation is done using Google Bi T-S model [35] pretrained on Image Net-1k with Res Netv2-101 [20]. We also evaluate the performance on Squeeze Net [29]... and on T2T-Vi T-24 [67]. Rank Feat performs the fusion at the logit space and computes the score function as log P exp((y +y )/2). The test batch size is set as 16. The approximate solution by PI yields competitive performances. (Table 7 shows results for PI with #100 iter, #50 iter, #20 iter, #10 iter, #5 iter)