A Primitive Based Machine Learning Framework for the Network Data-plane

The SDN-based network architecture currently lacks a framework for the efficient development and deployment of machine learning (ML) functions within the data plane. This paper addresses this gap by proposing a unified, modular ML framework designed to support data-plane deployment. The framework operates by identifying a minimal set of key primitives that can be effectively shared across various ML models. We demonstrate its functionality using defined primitives for Artificial Neural Network (ANN) models, incorporating specific approximations to enhance computational speed. We benchmark the framework through simulations to evaluate approximate resource usage and cycle computation. Our results indicate that these approximations result in only a minimal accuracy loss.