Artificial intelligence is tragically misnamedArtificial intelligence is a really, really bad name for this technological domain. This name, and what it implies, is responsible for so much of the misunderstanding. The concept of AI is, always has been, and always will be associated with the ubiquitous science fiction story of robots who compete against humans — and possibly replace them. The minute you describe any technological feat as AI, it inherits all that sense of excitement, fear, and attaches to itself many misleading interpretations.
The first misunderstanding is that there’s a fundamental difference between what we’ve achieved so far in computer science and what’s happening now and being branded AI. There are differences to be sure in every technological improvement but there’s no clear distinction between typical algorithms and so-called learning algorithms. Computer science, like any other field of science, has an ever-moving frontier separating the solved problem from unsolved problems. Things in the solved problem category, we might describe as things that can be automated and handled by an algorithm. Other things are those we can’t yet automate. Only humans can do those things. We say they require “intelligence”. This leads us to a good and only slightly tongue-in-cheek definition of the field of AI.
It’s not about creating intelligent machines which we can instead regard as an oxymoron. Any intelligence associated with a program or product is the intelligence of its human designers. Therefore, there’s nothing artificial about it.
The second important thing to understand is that AI, this field of automating more and more of what we do, is a process of building better tools to make us more productive, which in turn makes us richer and (perhaps) happier. Whenever tools appear that change the way people work and raise productivity, there’s a sense of fear that this will allow employers to replace some — or even all — of the workers. This rarely happens. The reasons are economic in nature and apparently unintuitive. It’s based on the false idea of fixed demand for goods and services. In fact, richer, more productive people consume more of everything, creating more jobs everywhere else for everyone. One can also simply refer to the historical record that technology doesn’t eliminate work. There’s nothing different this time. It simply changes how we work.
Any argument over the meaning or definition of artificial intelligence will simply revert to one over the meaning of the term intelligence. But it’s an unproductive conversation and has little to do with what actually happens in so-called AI research. The conversations we want to be having are far more concrete. Can we create a program to play chess better than a human? Can we create an algorithm to identify objects in images? Can we develop a software program to automate the driving of an automobile on common roadways? These are all questions at least beginning to shift categories into the “Yes” column and so currently considered topics of artificial intelligence. It’s better to recognize this evolution as steady progress rather than some sudden and recent departure into some bold new era.
Areas of recent advance: pattern recognitionWhat areas of research are actually advancing along the cutting edge of AI? Everyone interested in AI today should read Marvin Minsky’s classic 1960 paper on AI where he classifies the types of problems we face in AI. Minsky describes four categories: search, pattern-recognition learning, planning, and induction. The major advances of late have been in the category of pattern-recognition learning. Early waves of AI were more focused on other categories.
A clear example of this can be seen with chess programs. Chess programs surpassed human chess players in 1997 around the time of the famous Kasparov versus Deep Blue match. Within just a few years they moved well beyond the best human chess players. The strategy utilized for this belongs in the Minsky category of search. In short, a computer can look at all chess moves for several moves into the future and pick the one that has the best result. This brute force search technique, aided by some clever heuristics and databases, is basically the technique used today for best-of-breed chess programs, such as Stockfish or Komodo.
However, recent advances in pattern recognition have allowed programs such as DeepMind’s AlphaZero to beat Stockfish (under certain constraints). What’s exciting about this is that the core program isn’t really written with chess in mind. It’s general enough to play many games, just as humans can play many different games. It simply needs to know the rules and how things are scored and plays millions of times versus itself. Soon it becomes a master player. In fact, it plays the game of Go better than the best human Go player. What seems to be interesting about Go is that the search technique has never been capable of competing with humans due to the much faster explosion in the number of move possibilities.
But this is just an example of a pattern recognition technique surpassing a search technique. It's not AI versus conventional computing. It's AI versus AI. Yet, what's exciting, besides simply winning, is its generality which translates into less developer effort required for similar applications in the future. That's the productivity enhancement of today's pattern recognition based AI.
Other clear advances are in the ability of programs to process images and classify them into objects just as humans can do. This is a famously difficult problem for algorithms, especially when applying pre-specified rules. There have been great advances over the past decade in this challenging field of computer vision.
The right view is that algorithms for pattern recognition are quite general in nature and can be used as the basis for solving many kinds of problems. Many of those problems are associated with perception and physical locomotion that are exhibited by humans and other animals. One might claim that we have enough algorithmic sophistication to emulate the behavior of simple animals like ants or bees. While that’s far from human level sophistication, an argument can be made that some things we do, even for a living, might not involve much more sophistication. Driving an automobile might be one of those activities; hence the exciting, investment and/or dread of the self-driving car. It was realized long ago that many tasks of manufacturing fall into that category and robotics has been thriving in manufacturing for decades, a trend which will continue.
Skepticism on the power of pattern recognition based AIIt’s an open question of whether driving an automobile will actually fall into the solved category any time soon. I have a lingering suspicion that some researchers are being overoptimistic. I believe they’ll soon learn that solving even 95% of the problem isn’t enough to form a profitable business around and retreat to more practical territory like collision avoidance. If that’s the case, it may not be true that millions of people employed as drivers will suddenly be out of work. If it takes 25 years to phase out human driving, no one will really notice. No one will be suddenly unemployed — just as the rise of the internet didn’t lead to sudden and massive layoffs of travel agents, librarians or telephone operators. Yet there will be some obvious benefits from lower ambition wins. A program to prevent people from changing lanes and colliding with someone in their blind spot isn’t difficult to create and provides enormous value.
While the near future employment of cab drivers is at least debatable, most of our jobs aren’t dominated by pattern recognition tasks and so seem secure. As tools for better pattern recognition become more common, we will sustain a productivity increase similar to the advances in search. We’ll spend less time wading through data looking for patterns. We will have to spend less time coming up with algorithmic rules for many business processes.
To see why, it’s useful to understand the kinds of cognitive tasks that today’s AI advances have little ability to handle. First, consider that games like chess and go. Despite being complex to play, they have very simple rules which are fixed. The problems we face in our jobs are very different in nature. Likewise, the data structures used to represent games of chess or images are extremely simple. Compare that to the data we ingest through visiting a lawyer and discussing a legal case or reading relevant legal material online. All the complexities of a lawsuit do not fit into a vector or array. And even if they did, the rules concerning legal maneuvering and the abstract goals are extremely difficult to define. The key technique of pattern recognition through brute force self-playing is in fact rather useless for such problems.
Consider the problem of putting a human on the moon. You cannot accomplish the task through pattern recognition and learning correlations between data elements. For example, humans need air. You don’t want an algorithm learning that through trial and error! Sophisticated human thinking is in fact largely dominated by the use of models involving causality. Lack of air causes death and which, in turn, causes mission failure. And we didn’t learn that knowledge in the space age. We learned it in other contexts and can apply it in many different settings. Lacking a causal framework limits pattern recognition and prevents technology based on it from really competing with humans at most high-level tasks.
The mathematics of causal reasoning has only been developed in the last couple of decades; much of it by the academic Judea Pearl. And it hasn't had nearly as much influence on machine learning as has observational statistics. Thus, machine learning of today is limited in much the same was sciences would be if they were entirely observational (such as astronomy). Experimentation, an interventional, not statistical technique, is required to compete with human reasoning for many, perhaps the majority, of tasks.
Causal learnings are better than statistical learnings in many ways. They are better "atoms" of knowledge with which to build larger, more complex chains of reasoning. "Lack of air" will result in mission failure independent of everything else going on around it. This is rarely true of pieces of statistical learning which seem eternally confined to the contextual cages in which they were created. It's not very portable knowledge, making it difficult to work with. This is a concept rarely understood among today's AI practitioners, let alone the technology community a whole. It's a stubborn fact that computing power and more clever architectures will not change. I expect some disillusionment to arise when practitioners begin to realize these inherent limits of pattern recognition learning, the current AI technique in vogue.