Benchmarking CIFAR-10 with Tsetlin Machines

Summary

This paper presents an optimised toolbox for image classification using Tsetlin Machines on the CIFAR-10 dataset, achieving a significant leap in accuracy. Tsetlin Machines, known for their energy efficiency and interpretability, have traditionally lagged in handling complex colour images like those in CIFAR-10. By integrating novel Tsetlin Machine Specialists and advanced image processing techniques, such as Canny edge detection, Histogram of Oriented Gradients, adaptive thresholding methods, and colour thermometers, this research enhances the Tsetlin Machines' performance.

A rigorous hyperparameter search yielded optimal configurations for these specialists, culminating in a new state-of-the-art TM accuracy of 82.8% on CIFAR-10. The paper highlights the scalability of the proposed TM Composites architecture, which allows multiple independently trained Tsetlin Machines to collaborate, effectively pooling their strengths to outperform previous benchmarks.

The study underscores the potential of Tsetlin Machines in image analysis, offering an alternative to deep learning models with fewer computational resources and better interpretability. This toolbox sets a new benchmark for using Tsetlin Machines on CIFAR-10, making it a foundation for future research into energy-efficient, interpretable machine learning models. Keywords: CIFAR-10, Tsetlin Machines, image processing, benchmarking, machine learning.

Read more

The full paper An Optimised Toolbox for Advanced Image Processing with Tsetlin Machine Composites is available from arxiv.