ETHEREAL - Energy-efficient and High-throughput Inference using Compressed Tsetlin Machine

Aug 27, 2026

The Tsetlin Machine (TM) provides a logic-based alternative to deep neural networks by learning propositional patterns from data literals via Tsetlin automata. In TM models, literals may appear in both positive and negative clauses, enabling their cancellation to be leveraged for compression. ETHEREAL introduces an iterative training method that excludes selected automata states to sparsify TM logic patterns while preserving highly class-correlated literals, maintaining strong accuracy. Compared with standard TMs, ETHEREAL reduces model size up to 87.54% with minor accuracy loss, and is validated on eight TinyML datasets on STM32F746G-DISCO against TM baselines, Random Forest, and Binarized Neural Network.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages7

Reading Time18 min

AccessFree to download

Introduction

Background: TM vs DNN

TM pipeline: Booleanization, Training, Inference

Motivation for compression and sparsity

Prior work and limitations (e.g., REDRESS)

ETHEREAL approach overview

ETHEREAL: Energy-efficient and High-throughput Inference using Compressed Tsetlin Machine Shengyu Duan Newcastle University Newcastle upon Tyne, UK [shengyu.duan@newcastle.ac.uk](shengyu.duan@newcastle.ac.uk) Rishad Shafik Newcastle University Newcastle upon Tyne, UK[rishad.shafik@newcastle.ac.uk](rishad.shafik@newcastle.ac.uk) Alex Yakovlev Newcastle University Newcastle upon Tyne, UK [alex.yakovlev@newcastle.ac.uk](alex.yakovlev@newcastle.ac.uk) arXiv :2502 .05640v 1 [ cs .LG] 8 Feb 2025 ABSTRACT The Tsetlin Machine (TM) is a novel alternative to deep neural networks (DNNs). Unlike DNNs, which rely on multi-path arithmetic operations, a TM learns propositional logic patterns from data literals using Tsetlin automata. This fundamental shift from arithmetic to logic underpinning makes TM suitable for empowering new applications with low-cost implementations. In TM, literals are often included by both positive and negative clauses within the same class, canceling out their impact on individual class definitions. This property can be exploited to develop compressed TM models, enabling energy-efficient and high-throughput inferences for machine learning (ML) applications. We introduce a training approach that incorporates excluded automata states to sparsify TM logic patterns in both positive and negative clauses. This exclusion is iterative, ensuring that highly class-correlated (and therefore significant) literals are retained in the compressed inference model, ETHEREAL, to maintain strong classification accuracy. Compared to standard TMs, ETHEREAL TM models can reduce model size by up to 87.54%, with only a minor accuracy compromise. We validate the impact of this compression on eight real-world Tiny machine learning (TinyML) datasets against standard TM, equivalent Random Forest (RF) and Binarized Neural Network (BNN) on the STM32F746G-DISCO platform. Our results show that ETHEREAL TM models achieve over an order of magnitude reduction in inference time (resulting in higher throughput) and energy consumption compared to BNNs, while maintaining a significantly smaller memory footprint compared to RFs. KEYWORDS Tsetlin Machine, Machine Learning, Model Compression, TinyML 1 INTRODUCTION The ever increasing demand for deploying machine learning (ML) in low-energy, resource-constrained edge applications presents a significant challenge for deep neural network (DNN) implementations due to their high computational demands. This has led to efforts to identify alternative low-complexity ML algorithms. One such alternative is the Tsetlin Machine (TM), which is a novel ML algorithm that has been demonstrated with lower complexity than DNN, while achieving comparable accuracy across a range of ML Permission to make digital or hard copies of part or all ofthis work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for third-party components of this work must be honored. For all other uses, contact the owner/author(s) . tinyML Research Symposium’25, February 2025, Austin, TX © 2025 Copyright held by the owner/author(s) . datasets and exhibiting inherent interpretability [9] . A TM marks a fundamental shift from DNN by relying primarily on logic operations, which for example could outperform a multi-layer neural network (NN) in terms of accuracy [20], while eliminating hundreds of thousands of multiply-accumulate operations. Figure 1 demonstrates a typical TM structure for supervised ML. The structure comprises three incremental processes: A. Booleanization: Before TM training and inference regimes, the input dataset is first expressed in the form of a set of literals, represented as Boolean data. These literals are derived through a data encoding process, known as Booleanization. Atypical Booleanization process uses fixed or dynami

FAQs

What is the key idea behind ETHEREAL compared with standard Tsetlin Machines?

ETHEREAL uses an iterative training strategy that excludes selected automata states to sparsify logic patterns while retaining highly class-correlated literals. This reduces model size and keeps classification accuracy largely intact.

How does ETHEREAL achieve compression in TM models?

It exploits how literals can appear in both positive and negative clauses, allowing cancellation effects to support compressed representations. ETHEREAL further reduces includes by sparsifying TM logic through automata-state exclusions in both polarities.

What performance improvements does ETHEREAL report on TinyML hardware?

On STM32F746G-DISCO, ETHEREAL TM models achieve over an order of magnitude reduction in inference time and energy consumption compared to Binarized Neural Networks, while using a significantly smaller memory footprint than Random Forest baselines.

Aug 27, 2026

Uploaded by Nguyễn Văn Học