Optimal Parameter and Neuron Pruning for Out-of-Distribution Detection
Authors: Chao Chen, Zhihang Fu, Kai Liu, Ze Chen, Mingyuan Tao, Jieping Ye
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments are performed on multiple OOD detection tasks and model architectures, showing that our proposed OPNP consistently outperforms the existing methods by a large margin. |
| Researcher Affiliation | Collaboration | Chao Chen1, Zhihang Fu1 , Kai Liu2, Ze Chen1, Mingyuan Tao1, Jieping Ye1 1Alibaba Cloud 2Zhejiang University {ercong.cc, zhihang.fzh, yejieping}@alibaba-inc.com |
| Pseudocode | Yes | Algorithm 1 OPNP Optimal Parameter and Neuron Pruning |
| Open Source Code | No | The paper mentions using the PyTorch library and refers to another repository for a baseline method (DICE), but does not provide a specific link or explicit statement for their own source code. |
| Open Datasets | Yes | Following prior works [43, 44], we utilize Image Net-1K, CIFAR-10 and CIFAR-100 as ID dataset. For Image Net-1K benchmark, 50000 test samples are used as test ID samples, and four datasets are used as test OOD data, which are from (subset of) i Naturalist, SUN, Place and Texture [44]. |
| Dataset Splits | Yes | The hyperparameters ρw min, ρw max, ρo min, ρo max are experimentally determined in the validation set, which includes 50000 test ID samples and 50000 test OOD samples that are selected from images21k. |
| Hardware Specification | No | The paper does not provide specific hardware details such as CPU or GPU models, or memory specifications used for running the experiments. It only mentions general use of deep models and libraries. |
| Software Dependencies | No | The paper states 'Implementation of this work is based on Pytorch library', but does not specify the version number for PyTorch or any other software dependencies. |
| Experiment Setup | Yes | The hyperparameters ρw min, ρw max, ρo min, ρo max are experimentally determined in the validation set, which includes 50000 test ID samples and 50000 test OOD samples that are selected from images21k. We utilize grid search to determine the optimal pruning percentage, where we vary ρw min = {0, 5, 10, 20, , 60}, ρw max = {0, 0.1, 0.3, 0.5, 1, 3, 5}, ρo min = {0, 5, 10, 20, , 50}, ρo max = {0, 0.5, 1, 5, 10, 20, , 50}. The same hyperparameters are adopted in the same model. For both CIFAR10 and CIFAR100 datasets, we train a standard Res Net50 [18] model for 100 epochs. The learning rate is set to 0.01 and cosine annealing [32] is utilized for learning rate decay. |