Dream the Impossible: Outlier Imagination with Diffusion Models

Authors: Xuefeng Du, Yiyou Sun, Jerry Zhu, Yixuan Li

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct comprehensive quantitative and qualitative studies to understand the efficacy of DREAM-OOD, and show that training with the samples generated by DREAM-OOD can benefit OOD detection performance. and 4 Experiments and Analysis In this section, we present empirical evidence to validate the effectiveness of our proposed outlier imagination framework. In what follows, we show that DREAM-OOD produces meaningful OOD images, and as a result, significantly improves OOD detection (Section 4.1) performance.
Researcher Affiliation Academia Xuefeng Du, Yiyou Sun, Xiaojin Zhu, Yixuan Li Department of Computer Sciences University of Wisconsin, Madison {xfdu,sunyiyou,jerryzhu,sharonli}@cs.wisc.edu
Pseudocode Yes Algorithm 1 DREAM-OOD: Outlier Imagination with Diffusion Models Input: In-distribution training data D = {(xi, yi)}n i=1, initial model parameters θ for learning the text-conditioned latent space, diffusion model. Output: Synthetic images xood. Phases: Phase 1: Learning the Text-conditioned Latent Space. Phase 2: Outlier Imagination via Text-Conditioned Latent. while Phase 1 do 1. Extract token embeddings T (y) of the ID label y Y. 2. Learn the text-conditioned latent representation space by Equation (2). end while Phase 2 do 1. Sample a set of outlier embeddings Vi in the low-likelihood region of the text-conditioned latent space as in Section 3.2. 2. Decode the outlier embeddings into the pixel-space OOD images via diffusion model by Equation (6). end
Open Source Code Yes Code is publicly available at https://github.com/deeplearning-wisc/dream-ood.
Open Datasets Yes Following [2], we use the CIFAR-100 and the large-scale IMAGENET dataset [8] as the ID training data. and A Details of datasets Image Net-100. We randomly sample 100 classes from IMAGENET-1K [8] to create IMAGENET-100. The dataset contains the following categories: n01498041, n01514859, n01582220, n01608432, n01616318, n01687978, ...
Dataset Splits No No explicit specification of train/validation/test dataset splits (percentages or counts) or citations to predefined splits for the ID training data (CIFAR-100, IMAGENET-100). While standard benchmarks often have predefined splits, the paper does not explicitly state which splits were used for their own training/validation process.
Hardware Specification Yes We run all experiments with Python 3.8.5 and Py Torch 1.13.1, using NVIDIA Ge Force RTX 2080Ti GPUs.
Software Dependencies Yes We run all experiments with Python 3.8.5 and Py Torch 1.13.1, using NVIDIA Ge Force RTX 2080Ti GPUs.
Experiment Setup Yes We train the model using stochastic gradient descent for 100 epochs with the cosine learning rate decay schedule, a momentum of 0.9, and a weight decay of 5e 4. The initial learning rate is set to 0.1 and the batch size is set to 160. and β is set to 1.0 for IMAGENET-100 and 2.5 for CIFAR-100. To learn the feature encoder hθ, we set the temperature t in Equation (2) to 0.1.