Machine learning on phones just got 1.6x more efficient — but only researchers will notice

A research team built a smarter way to train AI models directly on edge devices (phones, IoT hardware) by cutting the amount of data the model processes during training. The trick: instead of looking at every part of an image, the model focuses only on the parts that matter for the current task, and trains the prompt and classifier separately to reduce computation overhead.

Training AI models on edge devices is hard because phones and IoT hardware have almost no memory or power. This paper shows you can cut training time and energy use by about 40% while keeping accuracy nearly the same — but this only matters if you're actually trying to run continual learning on a phone, which almost nobody is doing yet. The real question is whether edge continual learning is a problem people actually need to solve, or if it's a capability researchers built before the market asked for it.