Studying physics can be very useful – even when it comes to machine learning. A digital ‘super-brain’ with built-in knowledge of the fundamental laws of nature can speed up the development of optical components for everything from quantum computers to eyeglass or camera lenses according to a new study from Chalmers University of Technology in Sweden.

“When we fed the super-brain information about the laws of physics, it immediately got much smarter. Our calculations now take one tenth of the time previously required,” says Philippe Tassin, professor at the Department of Physics and Astronomy, Chalmers University of Technology.

Speculative decoding is a technique used to speed up large language models (LLMs), but existing approaches were mostly developed for English and are less effective in other languages. Now, researchers from Japan have developed ADASPEC, a speculative decoding framework that automatically generates language-specific training data and adapts its vocabulary during inference. Tested across seven languages and seven tasks, ADASPEC consistently outperformed existing methods while remaining practical and efficient for real-world multilingual applications.

In spectroscopic analysis, the development of cost-effective handheld infrared spectrometers has revolutionized monitoring practices, particularly in industries like agriculture and food production. Analysis time and efficiency of processing the raw spectral data are further boosted by the implementation of machine learning models for identification and quantification purposes. However, the inherent spectral variations arising from the random spatial arrangement of inhomogeneous granular media pose accuracy challenges. To enable real-time monitoring with enhanced precision, a scanning approach is presented herein. It involves the utilization of microelectromechanical systems-based infrared spectral sensors with spot sizes of 3 to 20 mm, facilitating spectrospatial averaging by scanning over the sample surface. A model is developed for the absorbance repeatability of the nonhomogeneous samples. It takes into account the effects of spatial and temporal noise, the sample scanned area, and the optical throughput of the device, also revealing an optimum point for the light spot size. Through experimental validation performed on wheat, chosen among the major cereals, we demonstrate that the spatial scanning method considerably improves absorbance repeatability of the 10-mm sensor by up to 5 times compared to stationary measurements, showing a good agreement with the theoretical model. The impact of sample scanning on artificial intelligence-based chemometrics model accuracy is rigorously assessed using heterogeneous feed samples, revealing substantial enhancements in the accuracy of prediction models. The 10-mm sensor exhibits a substantial reduction in root mean square error for protein (−60%) and for moisture (−43.5%), 2 major nutritional quality indicators of cereal grains.

A study co-led by an SUTD researcher has developed an AI-powered deep-learning framework trained on experimental data rather than simulations, enabling faster and more accurate design of light-controlling nanostructures.

The goal of merging intelligent computers directly with the human body, whether for continuous health monitoring or controlling advanced prosthetics, has long been stalled by a fundamental physical conflict. A new review article in the International Journal of Extreme Manufacturing details how purely rigid architectures are shifted toward soft, brain-inspired electronics that can sense, store, and process information while mechanically conforming to biological tissues.