This past month, a software program designed to play an ancient game called Go defeated Lee Sedol, a South Korean gentleman who is an 18 time world champion, widely acknowledged to be the leading human player of the game.
The event didn’t attract much attention, probably because it was seen as a predictable, perhaps inevitable development. After all, computers have been capable of beating Chess Grand Masters for many years now.
However, if we pause for a moment and examine this a little more closely, we may find a deeper, more profound significance.
Go is a seemingly very simple game, hugely popular in the Far East. It’s played on a 19×19 grid of horizontal and vertical lines. Two players are provided with a bowl of either white or black stones. They take turns placing their stones on the points where the lines intersect. The object is to use your stones to claim territory on the board. Stones surrounded by enemy stones are considered “captured” and are removed from the board. The player with most stones on the board and territory wins. Play continues until somebody concedes.
The apparent simplicity of the game is actually quite deceptive. In chess, played on an 8×8 board with fairly restrictive rules as to how players can be moved, it’s been calculated that there are about 1047 different possible games that could play out. Although that’s a huge number, it is finite and therefore could be “solved” once computers developed sufficient processing power. Such programs are able to analyze any configuration of pieces and select options that will maximize likelihood of success based on an analysis of all possible outcomes.
The size of a Go board, and the simplicity of the rules mean that there are an enormous number of configurations and game possibilities. In fact, that number has been estimated to be 10170 (http://senseis.xmp.net/?NumberOfPossibleGoGames). That’s a number difficult to even conceive. To get some sense of its magnitude, let’s consider the following comparisons:
- Postulated time that has elapsed since the “Big Bang” (beginning of the universe) = 13.8 billion years = 4.335 x 1017
- Diameter of the observable universe = 93 billion light years = 8.8 x 1026
- Estimated number of atoms in the observable universe (according to Universe Today http://www.universetoday.com/36302/atoms-in-the-universe/): 1080.
So, suffice to say, 10170 is a pretty big number. In fact, it’s more of a concept than a number. Essentially, it’s infinity.
It’s difficult to understand what makes expert players succeed in a game so endlessly variable but, according to experts like Mr. Lee, it seems to be as much about creativity, spontaneous insights that emerge within a game, and much understanding of the tendencies of an opponent – all things we have considered to be uniquely human attributes.
What all this means is that the computer-based approach to the game must extend far beyond simply providing sufficient processing power to filter through possible outcomes. The computer has to develop what the programmers refer to as “intuition” developed through what they call (wait for this) “deep learning”.
Deep learning? From a computer?! Difficult for mere humans like myself to even grasp but it seems that, given enough processing power and enough historical game outcomes to review, the computer is able to analyze trends and resulting outcomes, eventually sorting through the “clutter” of countless individual human game experiences to develop principles, optimal approaches and even heuristic “rules of thumb”. In other words, it isn’t simply analyzing, it’s thinking.
Shortly after I’d read about Mr. Lee’s encounter with AlphaGo, I happened to overhear an interview on NPR between a rather enthusiastic computer programmer and somewhat bemused reporter. The programmer was making the case that the United States would be better off with an artificial intelligence President. In fact, he was making the case that this was inevitable within the next 15-20 years. Building on the success of AI approaches to complex games, he was making the case that a computer would be able to analyze all relevant facts, public opinions and historical events in coming to the most reasonable conclusion about any issue that might arise, and would do so without the various human frailties and inevitable personal/political influences that plague “human” political leaders. The interviewer, who seemed to initially approach the whole encounter in a humorous was, by the end, conceding that the AI “person” could at least serve as an impartial advisor to the human decision maker – for now.
All this raises some rather disturbing implications for the medical profession. Clearly, it’s not much of an extension to imagine artificial intelligence of this type finding its way to the development of “Artificial Doctors”. The ability to instantaneously consider all possible evidence, reference all prior outcomes and even “factor in” patient preferences without the nagging issues of personal distraction, fatigue or subconscious biases that plague mere humans seems hugely attractive, particularly when considering the emerging applications of robotic surgery and procedures. One can only imagine how governments and other funders who are struggling with the economic issues related to physician payment, might droll at the prospect of replacing physicians, or what they do, in this way.
This all begs the question, what will be the real value of physicians two or three decades into the future? Generations of physicians, to date, have earned their keep through their knowledge and technical expertise. As those commodities become available in alternative (and decidedly non-human) ways, what’s left? What “value” will mere human doctors provide? What implications does this have for the education we should be providing? I offer a few thoughts on this issue.
- It all starts with communication. Patients will always come as unique and diverse individuals with varying illness experiences. The ability to interpret their experiences in a way that will allow for the diagnostic and then treatment process to begin will always be rooted in a personal and human relationship.
- We will need more, not less basic science. A strong understanding in the underlying physiological and pathological processes that underlie disease and clinical presentations will continue to allow physicians to not only understand their patient problems, but also find creative and unique approaches to unusual or atypical presentations.
- Patients will always value the human interaction. Any study looking at what they value from their physicians prominently includes compassion and the personal interaction that they receive.
- Patients will need someone to advocate for them. Our health care system is complex, and this is likely to increase in the future. Patients will desire, and need, someone to help them navigate their illness experience. Our educational system should help medical students understand and learn how to utilize “the system” for the benefit of their patients.
Finally, it may all come down to a single word. “Care” is one of those interesting words that serves as both a noun and a verb. As such, it probably allows us a means to best describe the difference between artificial and human intelligence. Computer-based AI will, without question, be able to provide excellent, arguably superior, clinical decisions. However, they will never be capable of truly caring.
Vive la différence.
Anthony J. Sanfilippo, MD, FRCP(C)
Undergraduate Medical Education