Advancing Artificial Intelligence: An interview with Stuart Jonathan Russell, PhD
Clinical Innovators | Interview by Katlyn Nemani, MD
Stuart Jonathan Russell, PhD, is a professor of computer science and Smith-Zadeh Professor in Engineering at the University of California, Berkeley, CA. He also holds an appointment as adjunct professor of neurological surgery at the University of California, San Francisco, CA, where he researches computational physiology and intensive care unit monitoring. Dr. Russel is an expert in artificial intelligence (AI) and studies decision-making, probabilistic reasoning, learning, robotics, and the foundations of intelligent systems. He became a fellow of the American Association for the Advancement of Science in 2011, and, in 2012, he was appointed to the Blaise Pascal Chair in Paris. Dr. Russel is the author of Artificial Intelligence: A Modern Approach, a textbook used by over 1,300 universities in 116 countries.
What drew you to the field of AI?
Artificial intelligence studies the problem of intelligence and how it may be created in machines. It’s one of the most fundamental and difficult problems there is. Human intelligence is quite amazing, and nearly everything we have as humans comes from that intelligence. Having intelligent machines could extend our reach in much the same way as ordinary machines have extended our physical reach.
How is AI being used in the realm of medicine right now? What developments do you expect in the next decade?
Medical diagnosis is one of the oldest areas of AI research. Early work on rule-based systems for diagnosis showed promise, and Bayesian methods developed in the 1980s proved quite successful in combining diagnostic evidence according to the rules of probability theory to identify and evaluate the possible causal explanations for a patient’s symptoms. Unfortunately, it was very hard to integrate those systems into the typical physician’s workflow in those days; they asked a lot of questions, required a lot of typing, and were quite brittle because they did not have access to the “whole patient,” only to the circumscribed list of symptoms. In areas such as tissue pathology and mammography, where the evidence is the image and the image is now online, we are starting to see successful applications of AI technology; in cardiology, ECG interpretation has been semi-automated for quite a while. Robotic surgery is already a reality for certain procedures and will become more widespread as AI systems learn how to manipulate soft tissue and interpret visual and other imagery during a procedure. As electronic medical records, genomic information, and wearable monitors become ubiquitous, we’ll see useful AI systems, including personalized diet and health regimens, health alerts, long-term monitoring of chronic conditions, and long-term health cost prognosis for individuals. We are also seeing AI systems being used in research to interpret data, and to build and manipulate complex models of cells, tissues, and organs in order to understand the processes and develop new treatments.
What kind of work are you doing in the medical field?
For the last few years, I have been collaborating with Geoffrey Manley, MD, PhD’s group at UCSF. Dr. Manley is a leading expert in both neurotrauma and intensive care medicine. The goal is to apply AI methods to interpret the data collected from the patient in real time in the ICU. The approach involves building complex models of the underlying physiology—cardiovascular, pulmonary, intracranial, etc.—and of the associated sensor devices, and to combine those models with the measured data (using Bayesian techniques) to adapt the models to the individual patient’s physiology, and to assess the probabilities of pathophysiological states such as hematoma or cerebral vasospasm. Those probabilities are very useful for the nurse or physician in deciding on a course of treatment. Eventually, we’d like to contribute to the treatment-planning problem as well. Hundreds of medical procedures and dozens of drugs are administered in the ICU to a patient over an extended period, and, at the moment, it’s mostly reactive (just trying to keep the patient state within normal bounds). I’m hoping it’s possible to do a lot better and to reduce mortality, morbidity, and the enormous costs associated with intensive care.
You were asked to speak at the United Nations this past spring to address some of the dangerous applications of AI. Could you tell us about that?
The UN is very concerned about lethal autonomous weapons—robots that decide where to go and whom to kill. In April, I was asked to explain to the UN meeting in Geneva the basic concepts of AI and autonomy, and how they will be applied to autonomous weapons in the future. After considering this question for a while, I concluded that an arms race in this area would lead to cheap, mass-produced weapons of incredible agility and lethality that would leave humans largely defenseless. I helped write an open letter, promoting a treaty to ban fully autonomous weapons that has been signed by about 3,000 AI and robotics researchers.
Is superhuman AI a reachable goal?
I see no reason to suppose that progress in AI will come to a halt. The benefits of progress are potentially enormous, and so the rate of investment in research is likely to increase. Almost certainly the kinds of AI systems we build will not have much in common with human intelligence, any more than Google has much in common with a human librarian, so there will be no meaningful notion of “machine IQ” and no obvious cross-over point to superhuman AI. However, it seems likely that machines will exceed human capabilities in more and more spheres of activity, and that gradually those spheres will become more integrated, so that it will make sense to talk of general-purpose intelligent systems.
What constitutes a conscious machine?
Humans, and perhaps some animals, are conscious machines. We have really no idea how or why, and no idea how to make a conscious computer or to determine that a given computer is or is not conscious. Sorry!
What does the future of AI look like?
It depends on our choices. Artificial intelligence systems could provide the greatest increase in human capabilities and human happiness of any technological advance in history, mainly because they could enable so many other advances. However, this future requires solving an important open problem: ensuring that the objectives we put into the AI systems are perfectly aligned with those of humans, so that we are happy with the resulting behavior and we are confident that the AI system will always act as a faithful assistant. This is far from easy. We aren’t aligned with each other and sometimes not even with our own selves.
< Back to Listings