Stephen Nash discusses the mysteries of deep learning.
In a recent issue of The New York Times magazine, there was an article by Siddhartha Mukherjee entitled, “This Cat Sensed Death. What if Computers Could, Too?” Here’s a short passage:
"A 2-year-old black-and-white cat named Oscar was apparently better than most doctors at predicting when a terminally ill patient was about to die. The story appeared, astonishingly, in The New England Journal of Medicine in the summer of 2007. Adopted as a kitten by the medical staff, Oscar reigned over one floor of the Steere House nursing home in Rhode Island. When the cat would sniff the air, crane his neck and curl up next to a man or woman, it was a sure sign of impending demise. The doctors would call the families to come in for their last visit. Over the course of several years, the cat had curled up next to 50 patients. Every one of them died shortly thereafter."
No one knows how Oscar was able to do this.
The article goes on to describe a research project at Stanford University which is using “deep learning” to analyze the records of 160,000 patients to try to replicate Oscar’s ability. Deep learning, with its reliance on neural networks, isn’t as mysterious as Oscar the kitten, since it is based on algorithms written by people that can be fully examined and on data sets that can be analyzed in other ways. But the patterns that the neural network discovers may not match cleanly with our understanding of medicine and may be hard to describe in ways that make sense to humans, so there is a mystery there as well.
Recently, I have the sense that deep learning is being applied to every sort of problem, including by many researchers here in the Volgenau School. It has been a huge success in playing games such as Go and Chess. It is the backend for the more recent versions of Google Translate. It’s used for Siri. It’s used to analyze images. And on and on.
In many contexts, the use of deep learning has led to dramatic improvements in the ability of computers to solve problems. However, unlike older approaches that might have been based on rules and logic, the newer approaches are based on having the algorithms look for subtle patterns in vast amounts of data. The results sometimes seem more like our own gut instinct than on any sort of reasoning that we can explain.
I sometimes can observe my own brain doing something like this. I’m pretty good at arithmetic, so I’m often asked to work out the tip at the end of a restaurant meal. Even if someone else is handling the check, I’ll often take a quick glance. If there is a mistake, I’ll know that the answer is wrong long before I can understand what the mistake is. The numbers just don’t seem right. It’s as mysterious to me as it is to the people sitting around the table. There are some theories of how this works.
This reminded me of a book that I read years ago: The Taming of Chance by Ian Hacking, a study of the development of statistical ideas in the 19th century. Hacking writes that, at the end of the Napoleonic era, newspapers started publishing large amounts of data. People started to notice that, from year to year, there was astonishing regularity in, for example, the number of crimes or suicides. Actions that people considered to be under the control of individuals now seemed to be subject to the omniscient control of mysterious statistical laws. Resolving these mysteries was a magnificent achievement that spanned a couple of centuries.
Let us hope that the mysteries of deep learning lead us on a similar path.
Stephen Nash is Senior Associate Dean of the Volgenau School of Engineering. This column appeared in his weekly newsletter for the school.