Human Skill in Machine Learning
Predictive Modeling and Supervised Learning are staple techniques in the Data Science arsenal of algorithms. Supervised machine learning or predictive modeling covers a large set of techniques that all rely on representative historical data that includes both a (possibly very large) set of features and a dependent variable that is learned as a function of those features. Example applications of these tools include a wide variety of tasks: image classification (cats, street signs), sentiment analysis of documents, medical diagnosis, targeted advertising, music composition, automated translation, stock movement predictions, or games such as chess, Go, and poker. Today, machines have been shown to be equal or better at the above mentioned tasks than humans. In the recent past, there has been an ever expanding list of successes that now give rise to concerns that large parts of the human workforce might become superfluous, possibly even the very data scientists who traditionally labored hard to create these solutions. Despite all the recent advances, machine learning models are only as good as the skills and experience of the data scientists who created them. Indeed, as good as the technology and new algorithms are, more often than not, the challenges rest with the data. This talk takes on some of the more broadly asked questions around what machine learning is good at and what it is not. It covers a number of case studies where the skillful adjustments and alternative perspectives of a data scientist provide creative solutions.
Presenter: Claudia Perlich, Two Sigma