3 June 2021 | 11:00 (Online)

By: Detlef Nauck, Head of AI & Data Science Research, Applied Research Division, BT Group UK

What is an AI model that has been created by machine learning? Is it data or is it software? If we say an AI model is specified by its parameters then it looks like data. If we say an AI model processes new data to make decisions it looks like software. These views matter because they influence how we should manage an AI model throughout its lifecycle. When we build AI models in production we are facing a number of challenges. Is the data we want to use complete and free of errors or bias? Have we created meaningful features and do we understand how they have been created? Has the machine learning model been cross-validated correctly? Is the data we use for decision making of the same type and distribution than the data used for training? Are the decisions the model makes as expected or has bias or drift crept in? Do we continue to get the expected business value out of the model and when should we discard it? To answer these questions we need to comprehensively and continuously test an AI model and the data its uses. In this presentation we review what we can learn from test driven software development to transform machine learning into an engineering discipline.

In cooperation with the Jean Golding Institute, University of Bristol
More information: the Bristol Data Science Seminar Series