Splitting upload bandwidth by sensor importance boosts AI accuracy on edge devices.

Small devices like phones often collect data from multiple sensors, such as cameras and microphones. Training smart artificial intelligence models on these devices is hard because they lack the computing power to process all these data types at once. Split learning helps by running the first few model layers on the device and sending partial calculations, called smashed activations, to a central server. However, sending this data uses up precious network upload limits. Usually, systems compress all data types equally, wasting network space on inputs that matter less. The researchers created ModalShare to fix this. The central server checks how much each sensor stream actually helps the final prediction using a special score. Then, it tells the device to spend more upload budget on the most useful data. ModalShare boosted model accuracy significantly at five times compression without adding extra client work or uplink network traffic.

Why it matters

This helps devices with multiple sensors, like in phones or cars, train better AI models without needing faster internet. It means smarter apps can learn locally, saving data and improving performance.

Questions answered