It has been proclaimed all over media outlets that 2017 is the Year of Artificial Intelligence. I would bet we will hear that “this is the year of AI” for the next 10+ years! AI is bound to penetrate most industries and generate significant efficiencies all over the place, so that title is granted. Which brings me to the central question of today’s blog post:
By definition, artificial intelligence is intelligence exhibited by machines‒but that doesn’t get us anywhere. Let’s focus on defining artificial high intelligence, as nobody is worried about their electric pencil sharpener taking over the world.
So what is high intelligence? Well, that’s pretty complicated to pin down too. We will take the indirect approach and start by agreeing on what high intelligence isn’t.
A calculator multiplies two numbers faster than me. Does that mean it is smarter than I?
Nobody in their sane mind would think so, because both the calculator and I know exactly what to do with this problem‒the calculator is just faster. In other words, there is a known algorithm that both the calculator and I will use to find the solution of this problem, and the fact that the calculator can do math faster than me is a known fact‒but that’s all I’m willing to concede: machines can compute faster than me.
The above example pretty much describes the state of computing until the rise of AI. Indeed, the leading edge in computing power has been with supercomputers. Since the 1960’s these machines have held and broken all records of computing power year over year. Top supercomputers currently perform quadrillions of operations per second (Petaflops, 10^15 ops/sec), and because of that, they are able to complete highly complex algorithms very fast‒algorithms like those involved in weather forecasting, predicting potential asteroid impacts using gravitational models, help design airplanes, and so on. All these are very hard algorithms because of the amount of operations that are involved in solving them is massive, so supercomputers are needed. But in some sense, these impressive technological machines are not so different from a calculator.
In other words, while computing power is great, it does not equate to intelligence: the ability to follow through a complex algorithm does not equate to being highly smart. What about the ability to figure out an algorithm? That begins to sound right…
There is no known winning algorithm for playing poker so far (compare with tic-tac-toe), so the machine that beat the Poker Masters did so by being intelligent.
Indeed, Carnegie Mellon University didn’t give the AI a game-winning algorithm; rather, it gave her a training algorithm! Their code provides the machine with the ability to determine, by its own experiments, which playing strategies work better than others. From this, the machine chooses a strategy that it determines to be the best, and plays accordingly. This complex process is iterated many times as the game is played, and the strategies the AI chooses keep changing at all times. (This requires high compute power.) After 20 days the machine won the game against the Poker Masters, and the rest is history.
This is the current state of AI, and I believe that it constitutes a form of high intelligence. We are able to write code that helps powerful machines find winning strategies by themselves, and also give them the capability to self-correct on the fly. That’s pretty impressive, but not quite as impressive as human intelligence…
Beating the Poker Masters is a one-off feat, while deploying AI in healthcare is more of a never-ending process. Because of this, AI victories in healthcare tend to be less spectacular, but much more relevant.
So far, AI has played a significant role on some aspects in healthcare such as the discovery and classification of new types of cancer by analyzing microarray data. At Prognos we are currently interested in using AI for finding patterns that signal early onset of disease.
Are machines smarter than humans? Maybe they are better at playing some games, but I believe that they are not quite as intelligent as humans overall. AI has surpassed human capacity in specific, clearly demarcated tasks. When a machine is able to solve a problem humans haven’t solved, such as finding a proof of the Riemann hypothesis (a very hard math problem), and is able to explain her idea to researchers in the field, then the machine has achieved human-level intelligence. At the time AI teaches us and generates knowledge like we do, I’d be the first one to call her a colleague.
There is yet another level of intelligence beyond this one, feared by prominent scientists such as Stephen Hawking: there will be a time when a machine tries to explain something it has discovered, but no human is able to understand her. By then AI has surpassed human intelligence, and we are definitely at the edge of the so-called singularity. At that time I will go back to my electric pencil sharpener and unplug it.
Want to learn more? Leave us a comment below, or reach out to Fernando directly at firstname.lastname@example.org. Stay tuned for Fernando’s next blog post on AI in the healthcare industry!