Server calls reduced by 56%, processing speed up to 3.67 times faster

Lee Young-jun (left), a researcher at KAIST's School of Electrical Engineering, and Professor Lee Jae-gil of the School of Computing. [KAIST]
Lee Young-jun (left), a researcher at KAIST's School of Electrical Engineering, and Professor Lee Jae-gil of the School of Computing. [KAIST]

Researchers have developed a technology that allows smartphone AI to remember knowledge it has acquired and independently resolve similar problems without repeatedly consulting a remote server. Rather than sending every complex query to a high-performance server, the system reuses previously obtained knowledge, cutting server calls by more than half and boosting processing speed by up to 3.67 times.

KAIST announced Monday that a research team led by Professor Lee Jae-gil of the School of Computing has developed CURE — short for Cumulative Knowledge Reuse — a collaborative framework that efficiently links small on-device AI models with large server-side AI models on smartphones and other devices.

Small AI models embedded in smartphones face inherent limits in handling complex tasks due to restricted computing power and memory. To compensate, a common approach delegates only difficult queries to high-performance servers, but existing systems treat each server response as a one-time answer. Even when similar problems recur, the device must request server assistance every time, driving up data traffic and processing time.

The research team addressed this by accumulating knowledge obtained from the server directly on the device and reusing it for future queries.

CURE operates in three stages: the on-device AI first assesses whether it can solve a problem on its own, then draws on knowledge stored locally for harder tasks, and contacts the server only when locally stored knowledge proves insufficient.

The team validated the technology using vision-language models capable of processing images and text together. The on-device component used MobileCLIP2, while the server-side model used EVA-CLIP, which has about 18 billion parameters.

An AI-generated image illustrating CURE, a collaborative framework that efficiently links small on-device AI with large server-side AI. [KAIST]
An AI-generated image illustrating CURE, a collaborative framework that efficiently links small on-device AI with large server-side AI. [KAIST]

In experiments, CURE maintained accuracy close to that of processing all inputs through a large server model while reducing server calls by an average of 55.61 percent compared with conventional device-server collaboration methods.

Processing speed was also up to 2.80 times faster than existing collaboration approaches and up to 3.67 times faster than routing all inputs to the server.

The technology can be applied without retraining existing AI models. It requires only an additional storage space on the device to hold knowledge retrieved from the server, making it compatible with a wide range of AI models and services.

The research team expects the technology to expand beyond smartphone photo search and personalized AI assistants to robots and wearable devices. It could be particularly useful in environments with unstable connectivity, helping robots draw on accumulated knowledge to assess their surroundings independently. However, actual performance may vary depending on device specifications and network conditions, and further validation is needed.

"CURE is a technology that allows a small on-device AI to remember knowledge it has acquired once and reuse it, rather than repeatedly asking the server the same question," Professor Lee said. "It will serve as a foundation for efficiently deploying high-performance AI across a wide range of devices by reducing communication overhead and response time while maintaining high accuracy."


nbgkoo@heraldcorp.com