Information-Theoretic Diffusion

Authors: Xianghao Kong, Rob Brekelmans, Greg Ver Steeg

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

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate ITD’s efficacy by evaluating its discovered concepts on downstream tasks and through human studies. On ImageNet, ITD achieves an average accuracy of 75.3% when clustering features from a self-supervised model, significantly outperforming alternative approaches.
Researcher Affiliation Collaboration Mengyu Yang1, Yi-Fan Zhang2, Ethan Liu2, Daniel M. Blatman2, Gregory M. P. O’Hare3, Georgios Tzimiropoulos1, Jie Song2 1 University of Oulu, Finland 2 Huawei Noah’s Ark Lab, Ireland 3 University College Dublin, Ireland 4 Huawei Noah’s Ark Lab, China
Pseudocode Yes Algorithm 1 Information-Theoretic Diffusion (ITD)
Open Source Code No Our code will be publicly released upon acceptance.
Open Datasets Yes ImageNet-1K (ILSVRC 2012) is a widely used benchmark dataset for image classification tasks, containing 1.28 million training images and 50,000 validation images across 1,000 classes.
Dataset Splits Yes ImageNet-1K (ILSVRC 2012) is a widely used benchmark dataset for image classification tasks, containing 1.28 million training images and 50,000 validation images across 1,000 classes.
Hardware Specification Yes All experiments were run on NVIDIA A100 GPUs.
Software Dependencies No The paper mentions software like PyTorch, scikit-learn, Faiss, and Diffusers, but does not specify their version numbers.
Experiment Setup Yes The backbone feature extractor is ViT-B/16 from DINO [5], pre-trained on ImageNet-1K. We extract features from the penultimate layer of the backbone model. We use AdamW optimizer with a learning rate of 1e-4 and a batch size of 256. The diffusion process runs for 2000 steps with a linear noise schedule, and the number of clusters K is set to 1000 for ImageNet.