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

TL;DR

The paper tackles energy-efficient edge inference by compressing Tsetlin Machines using ETHEREAL, an iterative literal-exclusion training method. It demonstrates that excluding literals appearing in both positive and negative clauses can drastically reduce model size (up to 87.54%) with small accuracy loss, while delivering faster and more energy-efficient inference on STM32 hardware. The evaluation on eight TinyML datasets shows ETHEREAL outperforms or matches RF and BNN baselines in throughput and memory usage, highlighting a practical, interpretable compression strategy for TM in resource-constrained environments. This work advances edge AI by making logic-based TM inference scalable and efficient for real-world microcontroller deployments.

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.

Figures (10)

• Figure 1: The TM structure.
• Figure 2: TM feedback procedure, independently performed for each clause. For binary classification, $y$=1 or 0 suggests the sample belongs to the class or not, respectively; for multiclass classification and a TM for a certain class, $y$=1 or 0 suggests the sample belongs to the class or other classes, respectively.
• Figure 3: (a) Type I and (b) Type II feedback, where TA state remains unchanged for any other cases.
• Figure 4: (a) Accuracy and (b) model size during training, for MNIST, as an example. Accuracy tends to increase with more epochs, accompanied by an increment in number of includes.
• Figure 5: Number of includes for all complemented features, represented in a 28$\times$28 image coordinate, for MNIST digit '2', after (a) 1 and (b) 50 epochs. More includes are induced as training proceeds. The included features are located both around the digit outline (as relatively significant features) and near the image border (as less significant features).