Pareto-Optimal Bit Allocation for Collaborative Intelligence

In recent studies, collaborative intelligence (CI) has emerged as a promising framework for deployment of Artificial Intelligence (AI)-based services on mobile/edge devices. CI, the AI model (a deep neural network) is split between edge and cloud, intermediate features are sent from sub-model to cloud sub-model. this paper, we study bit allocation feature coding in multi-stream CI systems. We task distortion function rate using convex surfaces similar those found distortion-rate theory. Using such models, able provide closed-form solutions single-task systems scalarized multi-task Moreover, analytical characterization full Pareto set 2-stream k-task systems, bounds 3-stream 2-task Analytical results examined variety DNN models literature demonstrate wide applicability