Browse Teach Curate

Multi-Task Supervised Feature Compression for Split Computing

A single lightweight encoder that learns a compressed intermediate representation at its early layers, serving multiple vision tasks (classification, detection, segmentation) at once for split computing between a mobile device and an edge server. Trained by supervised compression with a rate-distortion-style objective, it lowers transmitted-feature size, end-to-end latency, and mobile energy consumption versus per-task baselines.

B9 Text

A Multi-task Supervised Compression Model for Split Computing Yoshitomo Matsubara *

Introduces Ladon, a multi-task-head supervised compression model for split computing -- the setting where the early layers of a deep network run on a weak mobile device and the rest run on an edge/cl…