CILAB Unveils Core On-Device Vision AI Technology at 'ICML 2026'

CILAB presented a research paper on lightweight vision artificial intelligence (AI) at the official workshop of 'ICML 2026', one of the world’s three major machine learning conferences.

According to CILAB on July 16, this research focused on addressing the 'infrastructure construction and operational costs' issue, which is both the biggest topic and barrier in the AI industry. Conventional large-scale AI models could only operate smoothly in server and data center environments equipped with expensive GPU hardware. In contrast, in real-world settings where AI is actually deployed—such as CCTV control centers, smart factories, and retail stores—lightweight 'on-device AI' technology that enables high-performance AI to run on low-spec devices is essential.

CILAB's research team noted that directly applying compression techniques—widely used as standard in language AI such as Large Language Models (LLMs)—to video AI results in performance degradation. The team identified that the cause was not an inherent flaw in the compression technique itself, but rather the 'learning criterion', and proposed a solution by switching to a knowledge distillation approach.

Experimental results showed remarkable performance recovery across all five video models analyzed. In particular, for small models designed for edge devices, recognition accuracy—which had dropped to 18.79% after compression—surged to 67.11%, recovering to about 92% of pre-compression performance, with no slowdown in processing speed.

With this technology development, companies are expected to drastically reduce infrastructure investment costs, as they can now run high-performance video analysis AI in real time on their existing on-site equipment without the need to purchase additional, expensive GPU servers. Through this, CILAB has demonstrated world-class technological excellence across all three key stages of the physical vision AI pipeline: ▲digital twin-based synthetic data generation, ▲Sim2Real learning, and ▲quantization and lightweight deployment.

Youjin Song, CTO at CILAB, stated, "The significance of this research lies in proving the practicality of on-device vision AI that maintains near-original model accuracy while drastically reducing model size without sacrificing speed," adding, "We will continue empowering our customers to immediately leverage high-performance video AI on-site without additional infrastructure investment, further enhancing our market competitiveness."

This content was produced with the assistance of AI translation services.