Researchers at the Korea Advanced Institute of Science and Technology (KAIST) have developed a new system that could make artificial intelligence on smartphones significantly faster while reducing reliance on cloud servers.
اضافة اعلان
The system is named "CURE", short for "Cumulative Knowledge Reuse." It allows a small AI model running directly on a smartphone to remember answers previously received from a larger server-side model and reuse this knowledge later. According to the researchers, the system successfully reduced server calls by 55.61% while maintaining an accuracy level close to sending all requests to the cloud.
How Does the System Reduce Server Reliance?
Smartphones typically rely on small, lightweight AI models due to limited processing power and memory. These models handle simple tasks quickly but can struggle with complex or unfamiliar images, prompting them to seek assistance from a larger model running on an external server.
In conventional systems, the server's response is treated as a one-time solution. If the phone encounters a similar situation later, it requests help from the server again even as the trend toward running more AI tasks directly on devices continues to expand.
CURE alters this mechanism through a three-stage process:
It first checks whether the phone can perform the task independently.
It searches through previously stored knowledge gained from prior server calls.
It connects to the server only if no suitable solution is found.
For example, if the phone fails to identify a car model in a photo, the server provides the answer once. The system then stores the core features of the car rather than saving the entire image, allowing the phone to recognize similar cars on its own when encountered again.
What Does This Mean for Future Devices?
During testing, researchers evaluated the CURE system using multimodal vision-language models. The "MobileCLIP2" model was run locally on the phone, while the server utilized a much larger model, "EVA-CLIP", featuring 18 billion parameters.
Tests incorporating network latency showed clear speed improvements: the CURE system operated up to 2.8 times faster than traditional device-server collaboration methods, and reached speeds up to 3.67 times faster compared to routing all tasks directly to the cloud.
The system does not require retraining existing AI models; instead, it simply adds a separate knowledge repository. This means it could potentially be deployed without replacing a smartphone's existing AI system entirely.
This approach may become increasingly significant as smartphones transition toward a more AI-reliant era. The same methodology could also be applied to robotics and wearable devices, particularly those with limited computing capacity or operating in environments with unstable network connectivity.
However, actual gains in speed and efficiency will vary depending on device specifications, storage capacity, network quality, and data types. While the technology requires further testing, it presents a simpler path toward developing faster, more efficient on-device AI.
Resource: Al Ghad.
