An AI technology that allows devices to store and reuse learned solutions has been developed. Researchers at KAIST have successfully reduced server calls by an average of 55.61% while maintaining accuracy by enabling small AI to utilize the knowledge of large server AI.
On September 21, KAIST announced that a research team led by Professor Lee Jae-gil from the Department of Computer Science developed a device-server collaboration technology called CURE (Cumulative Knowledge Reuse) that efficiently connects small AI on devices with large AI on servers.
Devices like smartphones primarily use lightweight AI models due to limitations in computational power and memory. While small models can quickly handle simple tasks, their accuracy diminishes with complex or unfamiliar images. Conversely, relying entirely on high-performance server AI improves accuracy but increases data transmission time, communication volume, and server processing costs.
This has led to research on a collaborative approach where the small AI on the device makes initial judgments and only sends difficult problems to the server. However, existing methods consumed server responses only once, requiring repeated requests for similar problems. The research team focused on enabling the server's answers to be stored as reusable knowledge for future use.
CURE operates in three steps: 1) the small AI determines if it can process the input itself, 2) if not, it utilizes stored server knowledge, and 3) it only requests assistance from the server when necessary. Instead of storing server responses individually, it summarizes shared characteristics and differences among multiple cases as knowledge. This allows the accumulated knowledge to be used not only for previously encountered cases but also for similar new cases.
The research team validated the performance using a vision-language model that understands both images and text. The small model for devices was MobileCLIP2, while the large server model was EVA-CLIP, which has 18 billion parameters.
Experimental results on various image classification datasets showed that CURE maintained accuracy comparable to processing all inputs through the large server model while reducing server calls by an average of 55.61% compared to existing device-server collaboration methods that do not accumulate knowledge. The processing speed was up to 2.8 times faster than previous collaboration methods and up to 3.67 times faster than sending all inputs to the server.
CURE can be applied without the need to retrain small and large models separately; it only requires adding a storage system for server knowledge to existing models. This technology can be utilized not only for smartphone image recognition and search but also for devices like robots and wearables that need to perceive their environment with limited resources. However, the actual effectiveness of the application may vary depending on device performance, storage capacity, communication environment, and input data characteristics, necessitating further validation across various devices and settings.
Professor Lee Jae-gil stated, “CURE is a technology that allows the small AI on devices to remember and reuse knowledge obtained from the server without repeatedly asking the same questions.”
This research was led by Dr. Lee Young-jun, a postdoctoral researcher at KAIST, with Professor Lee Jae-gil as the corresponding author. Co-authors include Dr. Kim Do-young from Amazon, Dr. Kang Jun-hyuk from LG AI Research, and Professor Song Hwan-jun from KAIST's Department of Industrial and Systems Engineering. The findings were presented on October 10 at the European Conference on Computer Vision (ECCV 2026) held in Malmö, Sweden, from October 8 to 12. This research was supported by the Ministry of Science and ICT and the National IT Industry Promotion Agency's project on 'Core Technology Development for Human-Centered AI' under the 'AI Learning Capability Improvement Technology Development' initiative.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.