Dataset Condensation with Differentiable Siamese Augmentation

Authors: Bo Zhao, Hakan Bilen

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on multiple image classification benchmarks demonstrate that the proposed method obtains substantial gains over the state-of-the-art, 7% improvements on CIFAR10 and CIFAR100 datasets. We show with only less than 1% data that our method achieves 99.6%, 94.9%, 88.5%, 71.5% relative performance on MNIST, Fashion MNIST, SVHN, CIFAR10 respectively.
Researcher Affiliation Academia 1School of Informatics, The University of Edinburgh, UK. Correspondence to: Bo Zhao <bo.zhao@ed.ac.uk>, Hakan Bilen <hbilen@ed.ac.uk>.
Pseudocode Yes Algorithm 1: Dataset condensation with differentiable Siamese augmentation. Input: Training set T
Open Source Code Yes 1The implementation is available at https://github. com/VICO-Uo E/Dataset Condensation.
Open Datasets Yes We evaluate our method on 5 image classification datasets, MNIST (Le Cun et al., 1990), Fashion MNIST (Xiao et al., 2017), SVHN (Netzer et al., 2011), CIFAR10 and CIFAR100 (Krizhevsky et al., 2009).
Dataset Splits No The paper provides the number of training and testing images for each dataset (e.g., '60,000 training and 10,000 testing images' for MNIST) but does not explicitly specify the size or percentage of a separate validation split or how it's used for hyperparameter tuning beyond general mentions of 'training and validation' in the introduction.
Hardware Specification No The paper does not provide specific details about the hardware used for its experiments, such as GPU models, CPU types, or cloud computing instance specifications.
Software Dependencies No The paper mentions components like 'Re Lu activation' and 'instance normalization' but does not specify version numbers for programming languages, deep learning frameworks (e.g., PyTorch, TensorFlow), or other key software libraries used for implementation.
Experiment Setup Yes We set K = 1000, ςS = 1, ηθ = 0.01, ηS = 0.1, T = 1/10/50 and ςθ = 1/50/10 for 1/10/50 image(s)/class learning respectively as in (Zhao et al., 2021). The minibatch sizes for both real and synthetic data are 256. The network parameters for all architectures are randomly initialized with Kaiming initialization (He et al., 2015).