Presentation Information
[B-6-63]Higher Compression for Delta Coding in
Split Computing via an Adaptive Range Coder
◎Hiroki Shiba1, Zhiyi Zhu1, Patrick Finnerty1, Chikara Ohta1 (1. Kobe University)
Keywords:
Split Computing,Intermediate Feature Compression,Delta Coding
Split Computing partitions a DNN between edge and cloud and transmits intermediate features for collaborative inference, where the transmission volume is the bottleneck. For video, adjacent frames' features are strongly correlated in time, so a closed-loop DPCM (delta coding) transmitting only the temporal difference from the previous reconstruction is effective. We focus on the encoder of this two-stage (quantizer + encoder) pipeline: the quantizer fixes the order-0 entropy H0 as a rate lower bound, while the encoder drives the achieved rate R toward H0. We replace the general-purpose compressor zlib with a lightweight order-0 adaptive range coder. Being lossless, it exactly preserves accuracy (top-1 agreement with the uncompressed-feature prediction) while improving only the compression ratio. On UCF101 with EfficientNetV2-S (split points SP3/SP4/SP6), it surpasses the prior method at all split and operating points, reducing actual bytes by about 7-32% (1.08-1.47x), with larger gains at higher compression and little dependence on the split point. Future work includes lossy encoders, quantizer redesign, context modeling, and latency evaluation on edge hardware.