MANTIS: Modulated Adaptive Neural Compression for Task-Informed Split Computing
Jan 1, 2026·
,·
1 min read
Jared Macshane
Et Al.
Abstract
MANTIS is a task-informed split-computing system for neural compression. A lightweight client-side task detector estimates task relevance, a conditional analysis transform injects that signal through Conditional Generalized Divisive Normalization, and task-specific entropy bottlenecks and edge heads decode compact task-shaped latents for UAV perception workloads.
Type
Publication
Under Review
MANTIS evaluates task-conditioned split compression on UAVid semantic segmentation, WAID wildlife detection, and Boreal Fire smoke detection. The system improves the low- and mid-rate task-utility frontier, reducing bitrate by up to 62.2% at matched downstream accuracy and improving average normalized task accuracy by up to 9.3% at matched bitrate relative to the best non-conditioned or static baseline at the matched comparison point.
