Data Augmentation with Diffusion for Open-Set Semi-Supervised Learning

Authors: Seonghyun Ban, Heesan Kong, Kee-Eung Kim

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

Reproducibility Variable Result LLM Response
Research Type Experimental Through a comprehensive suite of experiments, we show that our data augmentation approach significantly enhances the performance of SSL methods, especially in the presence of class distribution mismatch.To assess the effectiveness of DWD, we conduct experiments across a broad set of tasks in two settings.
Researcher Affiliation Academia Seonghyun Ban1*, Heesan Kong1*, Kee-Eung Kim1, 2 1Kim Jaechul Graduate School of AI, KAIST 2School of Computing, KAIST shban@ai.kaist.ac.kr, hskong@ai.kaist.ac.kr, kekim@kaist.ac.kr
Pseudocode Yes Algorithm 1 Discriminator-Weighted Diffusion (DWD) Training. Algorithm 2 DWD Image-Seeded Generation.
Open Source Code Yes We provide open source repository for reproduction.
Open Datasets Yes Our extensive experimental results, utilizing CIFAR-10, CIFAR-100 (Krizhevsky & Hinton, 2009), Image Net-30 (Deng et al., 2009), and Image Net-100 (Cao et al., 2022) datasets with six baseline methods
Dataset Splits No The paper mentions training and test sets and discusses "validation" in the context of model performance, but it does not explicitly provide the specific percentages or counts for a validation dataset split, nor does it cite a predefined standard validation split.
Hardware Specification Yes GPU : NVIDIA Ge Force RTX 3090 Ti, CPU : Intel(R) Core(TM) i9-10980XE
Software Dependencies No The paper mentions using the "official implementation of latent diffusion model (Rombach et al., 2022)" and the "Adam W" optimizer, but does not specify software versions for general dependencies like Python, PyTorch/TensorFlow, or CUDA.
Experiment Setup Yes For tasks associated with Cifar-10, the Wide Res Net-28-2 architecture Zagoruyko & Komodakis (2016)) was employed, with training conducted using the Adam W optimizer at an initial learning rate of 0.03 across 256 epochs 1,024 iterations per epoch.