CW Complex Hypothesis for Image Data

Authors: Yi Wang, Zhiren Wang

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

Reproducibility Variable Result LLM Response
Research Type Experimental We support the hypothesis by visualizing distributions of 2D families of synthetic image data, as well as by introducing a novel indicator function and testing it on natural image datasets.
Researcher Affiliation Academia 1Department of Mathematics, Johns Hopkins University 2Department of Mathematics, Pennsylvania State University.
Pseudocode No No pseudocode or algorithm blocks were found in the paper.
Open Source Code No We train three DDPM (Ho et al., 2020) models of the same architecture from the open source implementation (Vandegar, 2023) respectively on Shigh, Slow, and S. (This refers to a third-party implementation, not authors' own code release).
Open Datasets Yes Figures 3-5 compares Ik, at k = 100, for individual label classes in MNIST, FMNIST, SVHN, CIFAR-10, CIFAR-100 and Image Net...
Dataset Splits No In test, the classifier is 100% accurate on validation data from the Shigh and Slow, as well as on the separately generated data S high and S low. (This refers to validation data for a classifier used within the experiment, not for the main experimental setup or dataset splits to reproduce the primary findings.)
Hardware Specification No No specific hardware details (GPU/CPU models, memory, or specific computing environments) were mentioned for the experiments.
Software Dependencies No We train three DDPM (Ho et al., 2020) models of the same architecture from the open source implementation (Vandegar, 2023) respectively on Shigh, Slow, and S. (No specific version numbers for software dependencies are provided.)
Experiment Setup Yes We train three DDPM (Ho et al., 2020) models of the same architecture from the open source implementation (Vandegar, 2023) respectively on Shigh, Slow, and S. For fair comparison, number of SGD training steps is 50000 for all models, but batch size is doubled from 32 to 64 when training the model on S so that each image in Shigh is used approximately the same number of times when training on Shigh and S, and similar for Slow.