Literal Labs scores 50x better AI performance using ‘Tsetlin’ approach
Cette publication existe aussi en Français
UK startup Literal Labs has published benchmarks for anomaly detection using Tsetlin machine software that show an approximately 50x uplift in inference performance compared with the best alternatives.
Literal Labs measured AI performance on MLCommons’ MLPerf Tiny Benchmark using the ToyADMOS dataset, which focuses on identifying irregularities in audio data. Spotting anomalies in audio is often a fundamental part of identifying equipment that is in need of preventative maintainence.
The Tsetlin machine-based AI software model was run on an off-the-shelf ARM based 32-bit microcontroller and comparisons made against alternative AI software models on the same hardware. The result is that Literal Labs’ model outperforms the current state-of-the-art on the same hardware by using 52x less energy and producing accurate inference results 54x faster, the company said.
Literal Labs (Newcastle-upon-Tyne, England) is the trading name of Mignon Technologies Ltd. founded in 2023 by University of Newcastle academics Professor Alex Yakovlev and Professor Rishad Shafik. The company recruited veteran semiconductor executive Noel Hurley to lead the company earlier this year.
Former Arm executives join ‘Tsetlin machine’ startup
A Tsetlin machine is a type of pattern-learning automaton based on propositional logic. Such logical “machines” for learning behaviour are known in academia for their compactness compared with neural network approaches. It makes the approach a good fit for edge AI on cost-effective low-power MCUs, Literal Labs asserts.
Literal Labs builds software models using Tsetlin machines and voting algorithms combined with binarization and compression. In the case of the benchmark this achieved a code size of just 7.29kbytes, and this is the driver of the 50x performance uplift.
The company benchmarked its model using the NUCLEO-STMH7A3ZI-Q development board which features a 32-bit RISC ARM Cortex-M7 microcontroller.
“These results reflect our focus on solving industry-critical problems while minimising the cost outlay and environmental impact typical of AI-powered solutions,” said Leon Fedden, CTO of Literal Labs, in a statement. “This is game-changing for predictive maintenance in sectors ranging from manufacturing to logistics, where downtime is costly, and efficiency is paramount,” he added.
The company has published a whitepaper detailing the benchmarking that is available from its website.
Related links and articles:
News articles:
Vaire co-founders discuss adiabatic reversible computing
Arm engineering veteran joins ‘reversible computing’ startup
Infineon selects Polyn analog neural network for TPMS