How to scale neural models for enterprise-wide adoption.
Software products using Machine Learning (ML) have vast potential for businesses. In this wide realm we find neural machine translation models, for example, that can reduce translation times of texts, or natural language processing (NLP) algorithms, that can sort customer data in order to personalize offers. But harnessing Machine Learning's full value needs mature data processing workflows.
Setting up a solid data management process that translates recent advances in Machine Learning algorithms into software products which are easily accessible, maintainable and upgradable is not an easy task. Enterprises often fall into the Proof of Concept (PoC) trap, where projects stay in PoC stage without managing to mature to production. Only well-engineered products yield consistent economic benefit to enterprises at low costs.
In a global enterprise like BMW, translating text is an often necessary but time consuming and tedious task. Cutting down translation times from days to minutes with little to no manual work aids the business in working faster and more efficiently.
Over the past decade BMW’s translation department collected a significant volume of technical translations and developed a prescriptive multilingual dictionary that ensures consistent usage of terminology throughout the translation.
During a series of internal Proof of Concepts (PoC), the machine translation team at BMW created a methodology of injecting this data into the training process as well as enhancing the translation quality by using terminology during inference. This work resulted in trained neural models which are used in the translation pipeline.
A solution for enterprise-wide shared services
For BMW’s neural models to be used company-wide it was necessary to deploy them as a service for employees and other automated services.
Reply achieved this for its client by providing a low-cost shared service based on AWS for the entire enterprise using BMW’s state-of-the-art neural machine translation models with specific adaption to the automotive domain. First applications of the neural machine translation API have already been productionalized in 2020 revealing significant cost saving potential for the business departments.
The case at BMW is only one example of how the right data workflows can set up Machine Learning products for long term benefits. While the business cases can differ significantly, the architectural blueprint for the solutions remains very similar. Most of them are shaped as approaches based on Kubernetes clusters for training and inference.
Reply offers enterprises hands-on advice on how traditional development patterns and best practices can be combined for successful development. In this downloadable white paper, experts present concrete examples which illustrate the application of current Machine Learning life cycle techniques to fundamentally distinct use cases.
Data Reply is the Reply group company offering a broad range of advanced analytics and AI-powered data services. We operate across different industries and business functions, enabling them to achieve meaningful outcomes through effective use of data. We have strong competences in Big Data Engineering, Data Science and IPA; we build Big Data platforms and implement ML and AI models in a manner that is repeatable, efficient, scalable, simple and yet secure. We support companies in combinatorial optimization processes with Quantum Computing techniques that enable an engine with high computational performances.