LASER: Linear Compression in Wireless Distributed Optimization

Authors: Ashok Vardhan Makkuva, Marco Bondaschi, Thijs Vogels, Martin Jaggi, Hyeji Kim, Michael Gastpar

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

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically demonstrate the superiority of LASER over the baselines on the challenging tasks of (i) language modeling with GPT-2 WIKITEXT-103 and (ii, iii, iv) image classification on MNIST, CIFAR10 and CIFAR100. With high gradient compression (165 ), LASER achieves 5064% perplexity improvement in the low and moderate power regimes on WIKITEXT-103.
Researcher Affiliation Academia Ashok Vardhan Makkuva * 1 Marco Bondaschi * 1 Thijs Vogels 1 Martin Jaggi 1 Hyeji Kim 2 Michael Gastpar 1... 1School of Computer and Communication Sciences, EPFL, Lausanne, Switzerland 2Department of Electrical and Computer Engineering, UT Austin, Austin, TX, USA.
Pseudocode Yes Algorithm 1 LASER
Open Source Code Yes Code is available at https: //github.com/Bond1995/LASER.
Open Datasets Yes We empirically demonstrate the superiority of LASER over the baselines on the challenging tasks of (i) language modeling with GPT-2 WIKITEXT-103 and (ii, iii, iv) image classification on MNIST, CIFAR10 and CIFAR100.
Dataset Splits No The paper uses standard benchmark datasets like WIKITEXT-103, CIFAR10, CIFAR100, and MNIST, but does not explicitly state the training, validation, and test splits (e.g., percentages or exact sample counts) within the provided text.
Hardware Specification No No specific hardware details such as GPU models, CPU models, or memory specifications were mentioned for running the experiments.
Software Dependencies No The paper mentions models and optimizers like 'GPT-2 like architecture' and 'Adam W', but does not provide specific version numbers for software dependencies or libraries (e.g., PyTorch, TensorFlow, Python versions).
Experiment Setup Yes Table 6: Default experimental settings for the GPT-2 model used to learn the WIKITEXT-103 task. ... Table 8: Default experimental settings for the RESNET18 model used to learn the CIFAR10 task.