Category-Extensible Out-of-Distribution Detection via Hierarchical Context Descriptions

Authors: Kai Liu, Zhihang Fu, Chao Chen, Sheng Jin, Ze Chen, Mingyuan Tao, Rongxin Jiang, Jieping Ye

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct extensive experiments to show the proposed hierarchical context descriptions are crucial to precisely and universally define each category. As a result, our method consistently outperforms the competitors on the large-scale OOD datasets, while showing comparable or even better generalization than the remarkable zero-shot methods. In this section, we empirically validate the effectiveness of our CATEX on real-word large-scale classification and OOD detection tasks.
Researcher Affiliation Collaboration 1Zhejiang University, 2Alibaba Cloud
Pseudocode No The paper describes its methods in prose and figures, but does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code No The paper does not provide a direct link to its own source code or explicitly state that its implementation code is publicly available.
Open Datasets Yes Datasets. Following the common benchmarks in the literature [59, 50, 60, 38], we mainly consider the large-scale Image Net [11] as the in-distribution data. Subsets of i Naturalist [53], SUN [65], Places [69], and Texture [8] are adopted as the OOD datasets.
Dataset Splits Yes For pre-processing, we follow Ridnik et al [42] to clean invalid classes, allocating 50 images per class for validation, and crop-resizing all the images to 224 resolution.
Hardware Specification Yes We use Python 3.7.13 and Py Torch 1.8.1, and 2 NVIDIA V100-32G GPUs.
Software Dependencies Yes We use Python 3.7.13 and Py Torch 1.8.1, and 2 NVIDIA V100-32G GPUs.
Experiment Setup Yes Following the default setting [71], each context consists of 16 learnable 512-D prompt embeddings, which are trained for 50 epochs using the SGD optimizer with a momentum of 0.9. The initial learning rate is 0.002, which is decayed by the cosine annealing rule.