This technique reduces memory use by 90% on small devices without losing accuracy.

Small devices like microcontrollers do not have the strong computer chips needed to run modern artificial intelligence. Today, methods that try to fit these AI models onto tiny devices either run very slowly or make too many mistakes. Researchers created GaLe to run already trained models without any retraining. GaLe splits the model's internal data into two parts. One part keeps exact, fine details for small local areas. The other part keeps a rough, overall picture across the whole image to track distant connections. This design lets small devices handle complex tasks that look across entire images. Tested on the ImageNet dataset, GaLe matched the accuracy of full-size models while cutting memory use by up to 90% and running up to 65% faster.

Why it matters

Embedded devices like wearables and sensors often can't run powerful AI models because they lack memory and processing power. GaLe makes it possible to use these models without retraining, so devices can be smarter while using fewer resources. This matters for developers building apps for IoT, healthcare gadgets, or any low-power device.

Questions answered