Increasing the Utility of Machine Learning in Clinical Care & On the Equivalence of Simulated Annealing & Interior Point Path Following for Optimization
Add to Google Calendar
Jenna Wiens: In recent years, the availability of clinically relevant datasets has rapidly grown. This has led to the investigation of machine learning (ML) techniques for improving clinical decision making. However, for data-driven systems to become widely and safely adopted in clinical care, there remain several key research challenges that the ML community must address: i) poor adaptability, current solutions do not readily adapt to complex changes in clinical data, ii) insufficient intelligibility, existing tradeoffs between accuracy and interpretability often result in highly accurate but uninterpretable models, and iii) absence of actionability, the present focus is on accurate predictions rather than on both accurate and actionable predictions. In this talk, I'll present/motivate these challenges in the context several clinical applications and discuss new and ongoing research directions in ML that aim to tackle these issues.
Jacob Abernethy: A well-studied deterministic algorithmic technique for convex optimization is the class of so-called "interior point methods" of Nesterov and Nemirovski, which involve taking a sequence of Newton steps along the "central path" towards the optimum. An alternative randomized method, known as simulated annealing, involves performing a random walk around the set while "cooling" the stationary distribution towards the optimum. We will show that these two methods are, in a certain sense, fully equivalent: both techniques can be viewed as different types of path following. This equivalence allows us to get an improved state-of-the-art rate for simulated annealing, and provides a new understanding of random walk methods using barrier functions.
Jenna Wiens: Jenna Wiens is an Assistant Professor in Computer Science and Engineering (CSE) at the University of Michigan. Her primary research interests lie at the intersection of machine learning and healthcare. The overarching goal of her research agenda is to develop the computational methods needed to help organize, process, and transform clinical data into actionable knowledge.
Jacob Abernethy: Prof. Jacob Abernathy is currently an Assistant Professor in the Department of Electrical Engineering and Computer Science at the University of Michigan. In October 2011 he finished a PhD in the Division of Computer Science at the University of California at Berkeley, and then spent nearly two years as a Simons postdoctoral fellow at the CIS department at the University of Pennsylvania, working with Michael Kearns. Abernathy's primary interest is in Machine Learning, with a particular focus in sequential decision making, online learning, online algorithms and adversarial learning models. He did his Master's degree at TTI-C, and his Bachelor's Degree at MIT. Abernathy's thesis advisor was Peter Bartlett.