AI Memory Stored on Devices Lowers Reliance on Servers

According to AI TIMES, a KAIST research team said on September 21 that it developed a collaborative technology called "CURE" that lets a smartphone’s small AI store and reuse judgments obtained from a server inside the device, cutting server calls by an average of 55.61%. The key feature of the technology is that it maintained high accuracy while also boosting processing speed by up to 2.80 times compared with existing collaborative methods. It works by having the device AI make a decision on its own first, then check stored knowledge if needed, and ask the server only when that is still not enough. Unlike conventional methods, it reuses answers received once for similar problems instead of consuming them each time. The researchers said it was designed to retain only the features needed for classification rather than storing the entire image, allowing it to respond to new, similar inputs as well. Because it can be attached without retraining, the team also suggested its potential use not only in mobile recognition but also in devices in environments with unstable connectivity.