Design of a Cloud-Based Data Platform for Standardized Machine Learning Workflows with Applications to Transport Infrastructure

Abstract

This paper describes the IBM One-Click Learning (OCL) platform, an ongoing internal effort in IBM Research aimed at providing end-to-end support for the full data science process. The design decisions behind the platform are presented, specifically with regard to promoting adoption and applicability within application domains where the potential of machine learning (ML) is recognized but still not fully harnessed. We focus on the representative case study of applying ML to civil engineering, a domain of growing interest also thanks to the EU coordination and support action to improve the European standards for inspection, monitoring and maintenance of bridges, tunnels and other types of transport infrastructures. This example is illustrative of the major roadblocks for adoption of ML in new application domains: the diversity in user profiles and familiarity with data science among domain practitioners; the variety in available hardware infrastructure and computing needs; and the heterogeneity, specificity and unsettled evolution of the use case landscape. Removing these roadblocks requires designing a platform explicitly geared towards usability goals of broad accessibility on one hand, and extensibility and specialization on the other one. We elaborate on a set of functionalities to support these usability goals and thereby enable the design of a task agnostic end-to-end platform to drive ML adoption and workflow standardization in new dynamic application domains. Finally, we present the IBM OCL platform as a proof-of-concept implementation of these functionalities and validate it in a use case where computer vision models are deployed to aid visual inspection of a bridge.