Out-Of-Distribution Detection with Diversification (Provably)
Authors: Haiyun Yao, Zongbo Han, Huazhu Fu, Xi Peng, Qinghua Hu, Changqing Zhang
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show that diverse Mix achieves superior performance on commonly used and recent challenging large-scale benchmarks, which further confirm the importance of the diversity of auxiliary outliers. |
| Researcher Affiliation | Academia | Haiyun Yao1, Zongbo Han1, Huazhu Fu2, Xi Peng3, Qinghua Hu1, Changqing Zhang1 College of Intelligence and Computing, Tianjin University1 Institute of High Performance Computing, A*STAR2 College of Computer Science, Sichuan University3 |
| Pseudocode | Yes | The whole pseudo code of the proposed method is shown in Alg. 1. |
| Open Source Code | Yes | Our code is available at https://github.com/Haiyun Yao/diverse Mix. |
| Open Datasets | Yes | ID datasets. Following the commonly used benchmark in OOD detection literature, we use CIFAR-10, CIFAR-100 and Image Net-200 as ID datasets. Auxiliary outlier datasets. For CIFAR experiments, the downsampled version of Image Net (Image Net-RC) is employed as auxiliary outliers. For Image Net-200 experiments, the remaining 800 categories from Image Net-1k (Image Net-800) serve as auxiliary outliers. |
| Dataset Splits | Yes | We can first determine the largest possible value of ω for the original baseline model while maintaining the ID classification accuracy. Then, we can select more suitable parameters for α and T, with adjustments made using an OOD validation set distinct from the testing OOD dataset. For example, a subset from the auxiliary outliers could serve as an OOD validation set. |
| Hardware Specification | Yes | We run all the experiments on NVIDIA Ge Force RTX 3090 GPU. |
| Software Dependencies | Yes | Our implementations are based on Ubuntu Linux 18.04 with Python 3.8. |
| Experiment Setup | Yes | We use Dense Net-101 [21] as the backbone for all methods, employing stochastic gradient descent with Nesterov momentum (momentum = 0.9) over 100 epochs. The initial learning rate of 0.1 decreases by a factor of 0.1 at 50, 75, and 90 epochs. Batch sizes are 128 for both ID data and OOD data. For Diverse Mix, we set α = 4, T = 10. |