Seoul National University unveils TACToC and mzCache to speed up mobile AI
Two new software frameworks cut wireless transmission costs up to 14 times and slash response delays on edge devices.
Engineers at Seoul National University have developed two software tools designed to reduce data traffic and latency when mobile devices offload tasks to remote artificial intelligence servers. The frameworks cut bandwidth consumption by up to 14.5 times and shorten initial smartphone response times by more than fivefold.
Selective transmission replaces video streaming for remote models
Inference offloading lets smartphones, robotics platforms and surveillance cameras send visual sensor data to external servers for complex artificial intelligence computation. Standard compression standards like H.265 prioritize visual fidelity for human eyes, transmitting unnecessary details that machine vision algorithms do not need. To resolve this inefficiency, a research team led by professor Kyunghan Lee at the Department of Electrical and Computer Engineering at Seoul National University created a technological innovation called TACToC, short for Task-Adaptive Communication for Task-oriented Communications, as reported by TechXplore.
A single encoder architecture cuts parameter switching to 12.5 milliseconds
Conventional task-oriented communication requires training dedicated encoders and decoders for every individual vision task. The TACToC architecture instead uses a single task-describable encoder combining relevant feature estimation with prompt-based task embedding, according to the research presented at the ACM MobiCom 2026 conference in Austin, Texas. The system estimates how much each visual feature contributes to an objective and selects only those features. Adapting the system to a new task requires updating roughly 1,900 parameters via lightweight trainable tokens rather than retraining the model, completing task switches in approximately 12.5 milliseconds.
Benchmark performance across five computer vision tasks
Testing across five standard workloads: semantic segmentation, pose estimation, image captioning, image classification and depth estimation, demonstrated communication cost reductions between 1.5 and 14.5 times while keeping accuracy loss within 3 percentage points. On an NVIDIA Jetson AGX Orin edge processing unit, the software processed approximately 21 frames per second. Under real-world wireless conditions, TACToC reduced bandwidth consumption by roughly 16.7% compared to H.265 video compression and by 41.1% compared to prior task-oriented communication schemes. The authors also introduced mzCache, a companion memory management framework that shortened the time to first response by up to 5.5 times when restoring machine learning applications on smartphones, marking a promising invention for any edge-computing startup.
Read also on The Patent: OpenAI and Anthropic claim AI solved a 27-year-old computer science problem
