Massively Parallel and Asynchronous Tsetlin Machine Architecture Supporting Almost Constant-Time Scaling

Authors: Kuruge Darshana Abeyrathna, Bimal Bhattarai, Morten Goodwin, Saeed Rahimi Gorji, Ole-Christoffer Granmo, Lei Jiao, Rupsa Saha, Rohan K. Yadav

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

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluated the proposed parallelization across diverse learning tasks and it turns out that our decentralized TM learning algorithm copes well with working on outdated data, resulting in no significant loss in learning accuracy. Furthermore, we show that the proposed approach provides up to 50 times faster learning. Finally, learning time is almost constant for reasonable clause amounts (employing from 20 to 7, 000 clauses on a Tesla V100 GPU).
Researcher Affiliation Academia 1Department of Information and Communication Technology, University of Agder, Grimstad, Norway.
Pseudocode Yes Algorithm 1 Decentralized updating of clause
Open Source Code Yes Code is available at https://github.com/cair/ICML-Massively-Parallel-and-Asynchronous-Tsetlin-Machine-Architecture.
Open Datasets Yes The Bike Sharing dataset contains a 2-year usage log of the bike sharing system Captial Bike Sharing (CBS) at Washington, D.C., USA. ... More details of the Bike Sharing dataset can be found in (Fanaee-T & Gama, 2014).
Dataset Splits Yes For both of the datasets, we use 80% of the samples for training and the rest for evaluation.
Hardware Specification Yes Our proposed architecture is implemented in CUDA and runs on a Tesla V100 GPU (grid size 208 and block size 128). The standard implementations run on an Intel Xeon Platinum 8168 CPU at 2.70 GHz.
Software Dependencies No The paper mentions software like CUDA, Thunder SVM, and pyTsetlinMachine, but it does not specify version numbers for these software components, which is necessary for reproducible setup.
Experiment Setup Yes We train on the known classes, which runs for 100 epochs with 5, 000 clauses, margin T of 100, and specificity s of 15.0. ... We train the TM for 100 epochs with a margin T of 500, 10, 000 clauses and specificity s = 25.0. ... The number of clauses n, margin T, and specificity s used are 300, 50, 5.0 respectively, for both datasets.