Dual Representation Learning for Out-of-distribution Detection
Authors: Zhilin Zhao, Longbing Cao
TMLR 2023 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments show that DRL outperforms the state-of-the-art methods for out-of-distribution detection. In this section, we demonstrate the effectiveness of the proposed DRL method. We compare DRL with post-hoc, confidence enhancement and ensemble methods. Furthermore, we analyze the effect of the hyperparameters and network backbones in DRL, run a set of ablation study experiments, and show the sensitivity of labeling information of labeland distribution-discriminative representations. |
| Researcher Affiliation | Academia | Zhilin Zhao EMAIL Data Science Lab, Macquarie University, Sydney, NSW 2109, Australia Longbing Cao EMAIL Data Science Lab, Macquarie University, Sydney, NSW 2109, Australia |
| Pseudocode | Yes | Algorithm 1 Dual Representation Learning (DRL) 1: Input: pretrained network gϕ, perturbation coefficient ϵ, covariance ΣD, batch size B 2: while no convergence do 3: Sample {(x1, y1), . . . , (x B, y B)} from P(X, Y) 4: Receive di = gϕ(xi), i [B] 5: Calculate ci = fθ(xi, di), i [B] 6: Estimate the objective function: i=1 log h(ci, yi) 7: Obtain gradients θ e L(θ) to update parameters θ 8: end while 9: Calculate out-of-distribution score: S(x) = max y [1,K] (h(c, y) + h(d, y)) /2 10: Output: S(x) |
| Open Source Code | Yes | 1The source codes are at: https://github.com/Lawliet-zzl/DRL |
| Open Datasets | Yes | We adopt CIFAR10 Krizhevsky (2009) and Mini-Imagenet Deng et al. (2009) as ID datasets to train neural networks. The numbers of classes of the two ID datasets are 10 and 100, respectively. We adopt CIFAR100 Krizhevsky (2009), CUB200 Wah et al. (2011), Stanford Dogs120 Khosla et al. (2011), Oxford Pets37 Parkhi et al. (2012), Oxfordflowers102 Nilsback & Zisserman (2006), Caltech256 Griffin et al. (2006), DTD47 Cimpoi et al. (2014), and COCO Lin et al. (2014) as OOD datasets to evaluate the OOD detection performance in the test phase. |
| Dataset Splits | Yes | We adopt CIFAR10 Krizhevsky (2009) and Mini-Imagenet Deng et al. (2009) as ID datasets to train neural networks. We adopt CIFAR100 Krizhevsky (2009), CUB200 Wah et al. (2011), Stanford Dogs120 Khosla et al. (2011), Oxford Pets37 Parkhi et al. (2012), Oxfordflowers102 Nilsback & Zisserman (2006), Caltech256 Griffin et al. (2006), DTD47 Cimpoi et al. (2014), and COCO Lin et al. (2014) as OOD datasets to evaluate the OOD detection performance in the test phase. |
| Hardware Specification | No | The paper does not explicitly mention the specific hardware used for running the experiments, such as GPU models or CPU specifications. It only states that the ResNet18 architecture was implemented in PyTorch. |
| Software Dependencies | No | The paper mentions implementing the networks in Py Torch but does not specify a version number for Py Torch or any other software dependencies. |
| Experiment Setup | Yes | We adopt the Res Net18 architecture He et al. (2016) for all the networks in the experiments and implement it in Py Torch. The learning rate starts at 0.1 and is divided by 10 after 100 and 150 epochs in the training phase, and all networks are trained for 200 epochs with 128 samples per mini-batch. If not specified, we set ϵ = 0.001 and adopt the same network architecture for pretrained and auxiliary networks in the proposed method. |