With the rise of large language models, two significant challenges are their high power consumption during training and their opaque nature. The Tsetlin machine (TM) offers a logic-based, interpretable alternative to traditional machine learning (ML) models to address these issues. TM is used in various ML-based edge computing and Internet of Things (IoT) applications, such as batteryless sensors, resource-constrained intrusion detection, and on-device training. It shows <inline-formula> <tex-math notation="LaTeX">$30.5\times $ </tex-math></inline-formula> reduction in computation cost, <inline-formula> <tex-math notation="LaTeX">$36.6\times $ </tex-math></inline-formula> reduction in storage memory footprint, <inline-formula> <tex-math notation="LaTeX">$13.5\times $ </tex-math></inline-formula> reductions in latency and energy, and converges faster than neural networks. It can be implemented on different hardware platforms, including field-programmable gate arrays, superconducting circuits, and memristor-transistor arrays. Filling a gap, we provide a holistic reference containing more than 160 implementations of TMs. In this tutorial and survey, we explain the algorithm of the basic TM in detail with a background of learning automata (LA). Next, we discuss different TM variants, including the regression TM (RTM), federated learning (FL)-based TMs, the coalesced TM (CoTM), real and integer-weighted TMs (IWTMs), and the convolutional TM (CTM) for text and image classification. We classify these software and hardware implementations based on their algorithmic similarities and compare them using key performance metrics such as accuracy, training time, energy consumption per classification, and operating frequency. In addition, we discuss our observations on general trends, insights from our comparative analysis, limitations of existing implementations, and potential directions for future research to further extend its applicability in energy-efficient IoT and edge computing applications.
Paper
Full text
A Comprehensive Review of Tsetlin Machines: Concepts, Applications, Analysis, and the Future
Semantic Scholar · 2026
Abstract
With the rise of large language models, two significant challenges are their high power consumption during training and their opaque nature. The Tsetlin machine (TM) offers a logic-based, interpretable alternative to traditional machine learning (ML) models to address these issues. TM is used in various ML-based edge computing and Internet of Things (IoT) applications, such as batteryless sensors, resource-constrained intrusion detection, and on-device training. It shows <inline-formula> <tex-math notation="LaTeX">$30.5\times $ </tex-math></inline-formula> reduction in computation cost, <inline-formula> <tex-math notation="LaTeX">$36.6\times $ </tex-math></inline-formula> reduction in storage memory footprint, <inline-formula> <tex-math notation="LaTeX">$13.5\times $ </tex-math></inline-formula> reductions in latency and energy, and converges faster than neural networks. It can be implemented on different hardware platforms, including field-programmable gate arrays, superconducting circuits, and memristor-transistor arrays. Filling a gap, we provide a holistic reference containing more than 160 implementations of TMs. In this tutorial and survey, we explain the algorithm of the basic TM in detail with a background of learning automata (LA). Next, we discuss different TM variants, including the regression TM (RTM), federated learning (FL)-based TMs, the coalesced TM (CoTM), real and integer-weighted TMs (IWTMs), and the convolutional TM (CTM) for text and image classification. We classify these software and hardware implementations based on their algorithmic similarities and compare them using key performance metrics such as accuracy, training time, energy consumption per classification, and operating frequency. In addition, we discuss our observations on general trends, insights from our comparative analysis, limitations of existing implementations, and potential directions for future research to further extend its applicability in energy-efficient IoT and edge computing applications.