MAPLE

Abstract

Many Artificial Intelligence (AI) applications are composed of multiple machine learning (ML) and deep learning (DL) models. Intelligent process automation (IPA) requires a combination (sequential or parallel) of models to complete an inference task. These models have unique resource requirements and hence exploring cost-efficient high performance deployment architecture especially on multiple clouds, is a challenge. We propose a high performance framework MAPLE, to support the building of applications using composable models. The MAPLE framework is an innovative system for AI applications to (1) recommend various model compositions (2) recommend appropriate system configuration based on the application's non-functional requirements (3) estimate the performance and cost of deployment on cloud for the chosen design.