Gradient-Based Novelty Detection Boosted by Self-Supervised Binary Classification

Authors: Jingbo Sun, Li Yang, Jiaxin Zhang, Frank Liu, Mahantesh Halappanavar, Deliang Fan, Yu Cao8370-8377

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In the evaluation with multiple datasets, such as CIFAR-10, CIFAR-100, SVHN and Tiny Image Net, the proposed approach consistently outperforms state-of-the-art supervised and unsupervised methods in the area under the receiver operating characteristic (AUROC) and area under the precision-recall curve (AUPR) metrics.
Researcher Affiliation Collaboration 1 Arizona State University 2 Oak Ridge National Laboratory 3 Pacific Northwest National Laboratory
Pseudocode Yes Algorithm 1: Gradient-based Novelty Detection Boosted by Self-supervised Binary Classification
Open Source Code No The paper does not provide any explicit statements or links indicating the availability of open-source code for the methodology described.
Open Datasets Yes Our experiment includes three in-distribution datasets: CIFAR-10, CIFAR-100 and SVHN. We test each of them using the other two as the OOD. In addition, we use two more OOD datasets: the resized version of the Image Net (Russakovsky et al. 2015) and LSUN (Yu et al. 2016) provided by (Liang, Li, and Srikant 2017).
Dataset Splits Yes For each IDD/OOD pair, we randomly select 5,000/5,000 (IDD/OOD) as the training dataset and further divide them into mini batches with size of 100/100 (IDD/OOD) to emulate the data streaming, the rest 5,000/5,000 (IDD/OOD) samples are used to test the binary classifier accuracy and the overall novelty detection performance after each new batch has been taken into the system.
Hardware Specification Yes All experiments are performed with Py Torch (Paszke et al. 2019) on one NVIDIA Ge Force RTX 2080 platform.
Software Dependencies No The paper mentions "Py Torch (Paszke et al. 2019)" but does not specify a version number for this or any other software dependency.
Experiment Setup Yes Our simple binary classifier has the structure of three convolution layers and one batch normalization layer with a Sigmoid classifier. It is trained by minimizing the cross-entropy loss using Adam (Kingma and Ba 2014). The initial learning rate is set to 0.0002 and the decay rate is controlled by β1 = 0.5 and β2 = 0.999. We train it for 500 epochs in both initial training and re-training process.