KAIST researchers develop CURE technology for efficient smartphone AI

A joint research team from KAIST, Amazon, and LG AI Research has developed ‘CURE’ (Cumulative Knowledge Reuse), a collaborative technology designed to connect small-scale AI on smart devices with large-scale AI on high-performance servers.

Existing collaborative AI methods often treat server responses as one-time use, requiring the device to request help from the server every time a similar complex problem arises. This increases communication traffic and latency. CURE addresses this by allowing the device to accumulate and reuse knowledge obtained from the server. The system follows a three-step process: the device's small AI first attempts to solve the problem, then attempts to use stored knowledge if the initial attempt fails, and only requests server assistance as a final resort.

In experimental tests using vision-language models, the CURE framework maintained accuracy levels close to processing all inputs via a large server model. However, it reduced server calls by an average of 55.61% and increased processing speeds by up to 3.67 times compared to traditional methods. The technology can be applied to existing AI models without retraining by simply adding a dedicated storage space for knowledge. Potential applications include smartphone photo searches, personal AI assistants, robotics, and wearable devices.