Synthetic data for model selection

Authors: Alon Shoshan, Nadav Bhonker, Igor Kviatkovsky, Matan Fintz, Gerard Medioni

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

Reproducibility Variable Result LLM Response
Research Type Experimental We perform extensive experiments on the CIFAR10 (Krizhevsky et al., 2009) dataset showing that indeed we are able to improve overall accuracy by selecting better models using synthetic data.
Researcher Affiliation Industry Alon Shoshan 1 Nadav Bhonker 1 Igor Kviatkovsky 1 Matan Fintz 1 G erard Medioni 1 1Amazon. Correspondence to: Alon Shoshan <alonshos@amazon.com>.
Pseudocode No The paper describes mathematical formulations and procedures like the linear ridge regression problem, but it does not present these in a structured pseudocode or algorithm block.
Open Source Code No The paper references third-party open-source projects like TIMM and pretrained-models.pytorch, and mentions that the generation code for Big GAN and DiT was publicly released by their respective authors, but it does not state that the code for the method described in *this* paper is open-source or available.
Open Datasets Yes In Sections 4.1 and 4.2, we perform experiments on CIFAR10 (Krizhevsky et al., 2009). ... In Section 4.3 we demonstrate rank preservation on Image Net (Deng et al., 2009).
Dataset Splits Yes In Sections 4.1 and 4.2, we perform experiments on CIFAR10 (Krizhevsky et al., 2009). In these sections, to evaluate the impact of the training set size, we use the following train-test splits: 10K-50K (Train10K), 30K-30K (Train30K), 50K-10K (Train50K). ... The standard protocol requires a validation set, to this end we split each of the datasets into training and validation subsets (train/val): Train10K was split into 7.5K/2.5K, Train30K was split into 22.5K/7.5K and Train50K was split into 40K/10K (see Appendix C for more details).
Hardware Specification No The paper does not provide specific details about the hardware (e.g., CPU, GPU models, or cloud computing instances) used for running the experiments.
Software Dependencies No The paper mentions software components such as StyleGAN2, WGAN-GP, DiT, BigGAN, TIMM, and PyTorch, but it does not specify their version numbers, which are necessary for reproducibility.
Experiment Setup Yes Appendix F (Standard Architectures Description) provides detailed hyperparameters for various deep model architectures, such as 'batch size 32, initial learning rate 0.05, depth 100, block type bottleneck, growth rate 12, compression rate 0.5' for DenseNet. Appendix G (Synthetic Data Generation Details (CIFAR10)) describes GAN training parameters, including how generators were selected based on FID scores.