Diffusion-based Layer-wise Semantic Reconstruction for Unsupervised Out-of-Distribution Detection

Authors: Ying Yang, De Cheng, Chaowei Fang, Yubiao Wang, Changzhe Jiao, Lechao Cheng, Nannan Wang, Xinbo Gao

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experimental results on multiple benchmarks built upon various datasets demonstrate that our method achieves state-of-the-art performance in terms of detection accuracy and speed.
Researcher Affiliation Academia 1Xidian University 2Hefei University of Technology 3Chongqing University of Posts and Telecommunications
Pseudocode Yes Algorithm 1 Training Algorithm; Algorithm 2 Testing Algorithm; Algorithm 3 Testing Algorithm for MSE Calculation; Algorithm 4 Testing Algorithm for LR Calculation
Open Source Code Yes Code is available at https://github.com/xbyym/DLSR.
Open Datasets Yes We train the OOD detection model on three in-distribution (ID) datasets: CIFAR-10 [Krizhevsky et al., 2009], CIFAR-100, and Celeb A [Liu et al., 2015].
Dataset Splits No The paper lists the ID datasets used for training (CIFAR-10, CIFAR-100, Celeb A) and OOD datasets for testing, but does not explicitly provide percentages or sample counts for training, validation, or test splits of the ID datasets used for their model's training.
Hardware Specification Yes Our method is trained on NVIDIA Geforce 4090 GPU for 150 epochs, with a batch size of 128 and a constant learning rate of 10 4 throughout the training phase.
Software Dependencies No The paper mentions specific components like "Efficient Net-b4", "Res Net50", and "Adam W optimizer", but does not provide specific version numbers for any software dependencies or libraries.
Experiment Setup Yes For Efficient Net-b4, we select feature maps from the first to fifth stages (M = 5) to construct the multi-layer semantic features, resulting in a feature dimension (c) of 720. The LFDN is consisting of 16 residual blocks. Inside each residual block, the number of groups in Groupnorm and the intermediate feature dimension of the residual branch are set to 1 and 1440, respectively. We employ the Adam W optimizer with a weight decay of 10 4. Our method is trained on NVIDIA Geforce 4090 GPU for 150 epochs, with a batch size of 128 and a constant learning rate of 10 4 throughout the training phase.