Out-of-Distribution Detection via Conditional Kernel Independence Model
Authors: Yu Wang, Jingjing Zou, Jingyang Lin, Qing Ling, Yingwei Pan, Ting Yao, Tao Mei
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical results demonstrate an evident performance boost on benchmarks against SOTA methods. We also provide valuable theoretical justifications that our training strategy is guaranteed to bound the error in the context of OOD detection. 5 Experiments In this section, we evaluate the performance of the Conditional-i method against the state-of-the-art OOD detection methods on both computer vision and NLP (natural language processing) applications. |
| Researcher Affiliation | Collaboration | Yu Wang 1 , Jingjing Zou 2, Jingyang Lin3, Qing Ling3, Yingwei Pan4, Ting Yao4, Tao Mei4 1: Qiyuan Lab, Beijing, China 2: University of California, San Diego, USA 3: Sun Yat-sen University, Guangzhou, China 4: JD AI Research, Beijing, China |
| Pseudocode | No | Not found. The paper describes the method textually and mathematically but does not include any pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code is available at: https://github.com/OODHSIC/conditional-i. |
| Open Datasets | Yes | In-distribution training data: We respectively use the CIFAR-100 and Image Net1K as the in-distribution training data for image OOD detection tasks. We employ the 20 Newsgroups data as the in-distribution training data for the NLP OOD detection task. By following the protocol in [28], we use the 80 Million Tiny Images2 [61] as the OOD training data when the In-distribution data is CIFAR-100 [32]. |
| Dataset Splits | Yes | In-distribution test data: We use the validation data from the corresponding in-distribution dataset as the in-distribution test data as defined in [28]. |
| Hardware Specification | Yes | We train each algorithm on 1 P40 GPU. |
| Software Dependencies | No | Not found. The paper mentions optimizers (SGD, Adam) and network architectures (Wide Res Net, Res Net18, GRU) but does not provide specific version numbers for software libraries or frameworks like Python, PyTorch, or TensorFlow. |
| Experiment Setup | Yes | We employ the SGD optimizer with lr = 0.1, Nesterov Momentum 0.9, for both Table 1 and 2. We use two-layer GRU [10] for producing Table 3 with Adam [31] optimizer. For Table 1 and 2, we adopt batchsize = 256 and 512 respectively for in-distribution data. For Table 3, we use batchsize = 64 for in-distribution data. Every benchmark method in Table 1 and 2 is trained for 100 epochs w.r.t. the size of the in-distribution training data. For NLP experiments, we train 2-layer GRUs [11] for 5 epochs for Table 3 |