A Scalable Architecture for Parallel Execution of the Tsetlin Machine

Abstract

With the Tsetlin Machine recently released, much research has been done on itscapabilities, with great success. However, the lack of tools available, and generalknowledge of the Tsetlin Machine prevents it from being adopted by the indus-try. As a result, it is today mostly used in academic environments. To increasethe general availability of the algorithm, this thesis introduces an introductorydescription to the algorithm and proposes an architecture that allows the use ofmultiple CPU threads and multiple GPUs to execute the algorithm in parallel. Inaddition, this thesis investigates several key aspects of the algorithm and how itcould be improved, like introducing dynamic self-learning parameters and parallelreduction techniques for the Tsetlin Machine. The results show that the pro-posed architecture improves execution speed. Further, the Tsetlin Machine wasable to adjust its own ”s” parameter from a bad initial parameter towards, andfinally converge with the known optimal value. The results from this thesis laythe foundation for creating powerful tools that allow for rapid development usingthe Tsetlin Machine in popular languages with high performance.