DRIVE: One-bit Distributed Mean Estimation

Authors: Shay Vargaftik, Ran Ben-Basat, Amit Portnoy, Gal Mendelson, Yaniv Ben-Itzhak, Michael Mitzenmacher

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

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our methods on a collection of distributed and federated learning tasks, using a variety of datasets, and show a consistent improvement over the state of the art.
Researcher Affiliation Collaboration Shay Vargaftik VMware Research shayv@vmware.com; Ran Ben Basat University College London r.benbasat@cs.ucl.ac.uk; Amit Portnoy Ben-Gurion University amitport@post.bgu.ac.il; Gal Mendelson Stanford University galmen@stanford.edu; Yaniv Ben-Itzhak VMware Research ybenitzhak@vmware.com; Michael Mitzenmacher Harvard University michaelm@eecs.harvard.edu
Pseudocode Yes The pseudocode of DRIVE appears in Algorithm 1.
Open Source Code Yes All the results presented in this paper are fully reproducible by our source code, available at [29].
Open Datasets Yes We use MNIST [51, 52], EMNIST [53], CIFAR-10 and CIFAR-100 [54] for image classification tasks; a next-character-prediction task using the Shakespeare dataset [55]; and a next-word-prediction task using the Stack Overflow dataset [56].
Dataset Splits Yes Detailed configuration information and additional results appear in Appendix E. We use code, client partitioning, models, hyperparameters, and validation metrics from the federated learning benchmark of [62].
Hardware Specification Yes Table 1: Empirical NMSE and average per-vector encoding time (in milliseconds, on an RTX 3090 GPU)...; ...using NVIDIA Ge Force GTX 1060 (6GB) GPU...
Software Dependencies No All the distributed tasks are implemented over Py Torch [45] and all the federated tasks are implemented over Tensor Flow Federated [46]. (Specific version numbers for PyTorch or TensorFlow Federated are not provided).
Experiment Setup Yes Detailed configuration information and additional results appear in Appendix E. We use code, client partitioning, models, hyperparameters, and validation metrics from the federated learning benchmark of [62].