Mark Schmidt is a Professor in the Department of Computer Science at the University of British Columbia. He has made significant impacts in the fields of both machine learning and numerical optimization.
Machine learning (ML) is a key tool we use to analyze the unprecedented amounts of data being collected in nearly all fields. ML is used in every day applications, such as automatic Zoom captioning, speech recognition on smart phones, face detection in cameras, product recommendations on websites, and car/pedestrian detection in cars. Advances in fundamental ML tools often lead to downstream impacts on many applications, and new applications will range from domains such as physics to biology and from education to human-computer interaction.
However, ML has a problem with what are known as hyper-parameters. The performance of modern ML models is very sensitive to these hyper-parameters. This situation will get much worse in the future as we use larger datasets, use models with more hyper-parameters to tune, and we use these models in more applications.
Schmidt’s work focuses on the development of substantially faster methods to deal with hyper-parameters. The key innovation of his work is adapting hyper-parameters at the same time as the model is learning from data. This has the potential to give similar performance to existing methods at a fraction of the cost.