Enhancing Inference Performance through Include only Literal Incorporation in Tsetlin Machine

The Tsetlin Machine is a novel and powerful algorithm for pattern recognition and decision-making tasks that has gained significant traction in recent years. Its features make it highly suitable for energy-efficient hardware implementations. This paper presents an FPGA design and implementation of an inference accelerator for a Multi Class Tsetlin Machine (MCTM). The proposed design aims at using the sparseness of the multi-class Tsetlin machine to optimize the inference algorithm implementation and provide a fast and resource-efficient implementation on Xilinx's Zedboard. We train the multi-class Tsetlin Machine model in software on the MNIST dataset and subsequently port the model to hardware. Further, we demonstrate and evaluate the performance of the proposed design using test images. It is observed that our design uses 21.1x less memory and is 30.8x faster compared to a standard design, with slightly more resource utilization for a given set of parameters.