The rule is, jam to-morrow and jam yesterday — but never jam to-day
I had the tremendous honor of being an attendee at the OurCrowd major yearly conference that took place earlier this week. Altogether, there were thousands of people there, all with very dynamic spirits that altogether combined, make up a good part of Israel’s startup nation. To be around so much energy and wisdom is humbling and inspiring.
I have the tremendous luck of being involved with one of the startups that was presenting its wares at the conference. Even just a few decades ago, who knows how many people took wondrous ideas with them to the grave. In today’s age, a person can develop an idea on their own or with a small group of friends, build a prototype, get the investors, and change the world. It still is very hard, very very hard, to get financial support for new ideas. I hope that over time there will be better tools available (even better than Kickstarter and Indiegogo) to help aspiring developers connect with supporters, so that very few good ideas fail to progress. Nevertheless, I found myself surrounded by people who had developed solutions for significant problems that, in some cases, I didn’t know existed.
Another measure of the success of the startup phenomenon is found in the magazines we read and the items we purchase. The classic example is the mobile phone. The iPhone, which really was the device to set in motion the worldwide spread of smart phones, is still not even 10 years old. I truly struggle to understand how we functioned before mobile communications and access to the Internet. But the release of the iPhone was just one flexion point along the exponential increase in the influence of technology in our day-to-day lives.
Just yesterday, I read about a new artificially intelligent [AI] system that managed to beat a top professional player in the game of Go. AI systems have been beating masters in various games for some time. The first time an AI system called Deep Blue beat a world champion in chess, was definitely perceived as a new era in computing. The issue though was that this chess AI system was based on a “brute force” approach. The AI basically tested for all possible future moves and their ultimate outcome. At a certain point, the system would decide which present move had the greatest likelihood of ending in checkmate against the human opponent. While this approach worked, it was clear to everyone that human minds did not work this way.
The game of Go is fundamentally different in that the number of possible moves at any point is far beyond the capability of any present day computer system to track. Therefore, the developers of this AI used a fundamentally different type of computer intelligence called deep learning. In many ways, deep learning mimics human learning. Human learning and understanding is based on a tremendous amount of pattern recognition, rather than hard calculation. Humans can look at a series of objects that they’ve never seen before, but still deduce that one of the objects is a dog and one of the objects is a cat and so on. Humans can do this because of our amazing ability to see a pattern and then compare it to stored patterns in our brains. In a fraction of a second, we have done all of the processing necessary to say that the animal we are looking at “looks like” a cat and not a dog. Deep learning simulates this type of pattern recognition. As it turns out, many complex problems can be solved by finding patterns rather than using brute force mathematics to try and come up with the solution.
The AI system that plays Go was able to beat a world-class master at the game. How was this achieved? Basically, the developers gave the AI a set of tens of millions of moves that expert human players had used. This sharing of real-life experience is in many ways similar to the way in which a parent teaches a child, a teacher educates a student and an artisan trains an apprentice. After this initial training of the AI, the developers had one version of the AI play against another version of the AI, and as such, the AI continued to learn from its own “real life” practice. Like humans, once they are taught something, they go and practice it on their own and thereby become proficient in it. The proof is in the pudding, and it is fair to say that this type of AI learning clearly works.
There should be no doubt that this is a flexion point further up on the exponential curve of advancing computer intelligence. The applications are immediately evident. Such a system can learn from the experience of every doctor throughout the world going back decades. Every case report is yet another piece of information that a medical AI could learn from. Eventually, you could have one medical AI present clinical cases to a second medical AI and the two could confer over the final diagnosis.
What is also fascinating about this whole story is the inaccuracy in predicting when this “AI wins at Go” achievement would occur. The developers themselves imagined that it would be many years, if not well more than a decade, before they would achieve significant success with the deep learning AI they had designed. They were themselves surprised by the quick success they achieved. This scenario is actually often repeated throughout the technology world. People are constantly making predictions about how long it will be before we live on Mars or cure cancer or end the occurrence of car accidents (via self driving cars, perhaps).
Ray Kurzweil, likely the most famous futurist in the world, is incredibly accurate in his predictions of when certain technologies will finally take hold. But for the rest of us, we often misread the signs and prophesize that future technologies will become part of our lives either much earlier or much later than in reality. How long will it be until medical AI systems are doing all of the interpretations of imaging studies that doctors rely on, on a daily basis? Will it be five years from now or 50 years from now? I personally believe that the number is somewhere in the middle. I also believe that there will be significant delays in the implementation of such medical AI systems because of the backwards thinking of many doctors, which literally delays medical technology advancement. Nevertheless, progress is inevitable and I would personally be surprised if such medical AI systems took more than 20 years to become standard.
Deep learning plays a big role in voice recognition and translation. Therefore, within this 20 years that I have imagined, I could very easily see a patient being initially interviewed by a medical AI, having all appropriate tests and initial treatments being ordered by the medical AI, and then only at the end, when the AI has basically made the diagnosis, the doctor is informed, [needlessly] reviews the chart and signs off. Once again, for very human reasons like fear and ego and small mindedness, it will probably be many decades before computers are truly allowed to treat patients on their own, in a manner that is better than any human can provide. But whenever this day arrives, the success of AI in playing the game Go, should be remembered as a milestone on the way towards totally automated medicine.
Many of the startups that presented at the OurCrowd conference could themselves become further flexion points on the exponential curve. The startup I am involved in, efficacy.care, is implementing a concept that is critical in medicine, yet has been greatly ignored. In many types of diseases, lifestyle changes are critical to the success of treatment. The importance of these lifestyle changes is reflected in the fact that they are referred to as the primary stage in the treatment of such diseases as high blood pressure, diabetes, high cholesterol, obesity and more.
However, there really is no mechanism for quantifying lifestyle, and as such, there is no way to formally measure whether a person is in fact improving in terms of lifestyle changes. Efficacy.care will use a series of tools, including deep learning, to create an index that will quantify lifestyle, so that it can be tracked in a far more exact way, in medical studies and daily medical management. I personally believe that this will become an essential part of the management of most patients. How long will it be until this type of lifestyle quantification is standard medical practice? I hope and believe that it will be quite soon. But maybe I should ask Ray what he thinks.
Thanks for listening.