Lunchtime Seminar - Monday 3 December - 12.30 Brownbag lunch, 13:00 Talk and Q&A
Location: CFI Boardroom, Level 1, 16 Mill Lane
Title: What Everybody Needs to Know About 'Prediction' in Machine Learning
Speaker: Momin M. Malik, Berkman Klein Center for Internet & Society at Harvard University
"Facebook ad feature claims to predict user's future behaviour"
"How your blood may predict your future health"
"Councils use 377,000 people's data in efforts to predict child abuse"
"Test could predict risk of future heart disease for just £40"
"MRI scan that can predict stroke risk has 'promise to save lives'"
"Fitbit could help doctors predict how patients will react to chemotherapy"
"Online test aims to predict best antidepressants for individual patients"
This is a sampling of article headlines in The Guardian from the past two years, all involving (whether explicitly mentioned or not) "computer models", "algorithms", or rather, machine learning. The articles capture hope and promise as well as terror: terror both of determinism, and the terror of who has knowledge of that determinism and who doesn't. But as we look closer, the narrative starts to break down. First, there are many references are to predicting "the future". What else would we predict, the past? Actually, yes. In "'I Wanted to Predict Elections with Twitter and all I got was this Lousy Paper': A Balanced Survey on Election Prediction using Twitter Data" (2013), Daniel Gayo-Avello writes, "It’s not prediction at all! I have not found a single paper predicting a future result. All of them claim that a prediction could have been made; i.e. they are post-hoc analysis and, needless to say, negative results are rare to find.” The technical meaning of a "prediction" in statistics and machine learning is of a "fitted value" of a model: that is, a value correlated with previous outcomes. This may be predictive, but there are many ways that this breaks down. While such usage pre-dates AI, going back to the early 20th century, its currency is perhaps an example of what AI researcher Drew McDermott critiqued in 1976 as "wishful mnemonics": AI researchers calling things what they want them to be, rather than describing what they are, by which they end up fooling themselves. With the attention put on AI, journalists and academics passing on this mnemonic as literal has likely contributed to widespread public misunderstanding and anxiety.
In this talk, I give what people researching philosophical, sociological, historical, ethical, and regulatory aspects of AI need to know about what "prediction" actually means, and how it works (and why it might not). Drawing from internal critiques in statistics and machine learning, I cover how 1) prediction is based entirely on correlations; spurious correlations can predict very well, but are highly vulnerable to changes in context; 2) methods to prevent overfitting, namely "cross-validation" with splitting data into "training" and "test" sets, can break down in the presence of dependencies, from the over-use of test data, and because of publication bias; and 3) the "bias-variance tradeoff" is a technical consequence of how predictive success is defined in which it is possible that a 'false' model predicts better than a 'true' model. By cover these topics, I aim to give researchers enough familiarity with technical concepts to critically approach claims made by and about machine learning and AI.
Biography: Momin M. Malik is the Data Science Postdoctoral Fellow at the Berkman Klein Center for Internet & Society at Harvard University. He received his PhD in Societal Computing and a master's in Machine Learning from Carnegie Mellon University's School of Computer Science. He also holds a bachelor's from Harvard University, where he studied the history of science, and a master's from the Oxford Internet Institute. His work uses data and modelling reflexively, performing in quantitative terms critiques from STS and other critical literature, as well as working with social scientists to both adopt and critique modern methods of data modelling.