This article originally appeared in the December 2016/January 2017 issue of WiFi HiFi Magazine.
Hollywood has given us many memorable visions of ‘artificial intelligence' (AI). HAL 9000 saying "I'm sorry Dave." Commander Data on Star Trek: The Next Generation wearing a Sherlock Holmes deerstalker hat. R2-D2 being comically pedantic. Runaway ‘hosts' on Westworld getting too smart for comfort. But lately, the term AI has taken on a very different meaning.
Work on actual machine intelligence hit a brick wall decades ago. We still have no idea what human intelligence is all about, so creating machines that can ‘think' like humans seemed about as likely as inventing faster-than-light space travel.
Nevertheless, many new products and services are being sprinkled with ‘AI' pixie dust, to make them sound more futuristic. "AI is the spice to make software delicious," said Jin Hyung Kim, Chairman of the Korea National Open Data Strategy Council, in a recent Microsoft-hosted panel on the subject.
The irony is that today's so-called ‘AI' systems don't need misleading nomenclature to make them more impressive. Even if they're not exactly ‘intelligent,' they do offer powerful, and eminently profitable, solutions to some very practical problems.
When we separate the hype from the fact, it turns out that there are two distinct AI stories to be told. One is a fascinating true story about advances in data processing. The other is an aspirational fantasy about the creation of a thinking machine.
Weak or Strong
Artificial Intelligence is a term originally coined by Dr. John McCarthy at Dartmouth College in 1956, referring to computers that could use language, form concepts, and generally solve the kind of problems that only humans have been good at. The Holy Grail would be a machine that could pass the test postulated by the great mathematician Alan Turing, indistinguishable from a human under cross-examination.
McCarthy eventually admitted that, "AI is harder than we thought." And indeed, today we still have nothing remotely resembling true AI, even in the most advanced laboratory. Early optimism started evaporating back in the 1980s, and many AI researchers moved on to simpler tasks that showed some hope of producing practical results.
What was initially meant by ‘AI' was unceremoniously renamed "strong AI." This sleight of hand has allowed companies like Microsoft and Google to operate huge projects in the field of ‘AI,' without tackling the intractable problems of actual machine intelligence. Strong AI "doesn't exist," sums up tech journal LifeHacker.com. "Artificial intelligence just means anything that's ‘smart.'"
Douglas Hofstadter, Professor of Cognitive Science at Indiana University and author of the Pulitzer-Prize-winning book Gödel, Escher, Bach, is one of relatively few researchers still committed to understanding the nature of human intelligence, and how it might be replicated by digital electronics. In a 2013 interview with The Atlantic, he referred to recent ‘AI' efforts as "trickery," and added: "I don't want to be involved in passing off some fancy program's behaviour for intelligence when I know that it has nothing to do with intelligence."
IBM's Watson became synonymous with artificial intelligence after it clobbered two human Jeopardy! champions by interpreting the quiz game's cryptic ‘answer' clues to complex questions using techniques like pre-coded logic rules and full machine learning.
On the other hand, dilettante inventor, entrepreneur and author Ray Kurzweil continues to strongly propound the opposing view: that true artificial intelligence is inevitable, and coming fast. While Hofstadter and others extrapolate conservatively, Kurzweil draws a curve shooting skyward towards The Singularity (in about 2045), when humans and computers will merge.
The divide is very wide. Apart from Hofstadter, Kurzweil's critics include linguist and MIT Professor Emeritus Noam Chomsky, and even Microsoft co-founder Paul Allen. They see Kurzweil's ideas as superficial, and not grounded in hard data from AI research.
Still, at a Singularity Summit a few years ago, Hofstadter willingly admitted the possibility that Kurzweil's predictions might prove valid. He asked only that they be debated and tested on a scientific rather than "hand-waving" basis.
Rote Learning
While the direct development of strong AI makes little obvious progress, a new type of computing technology has made rapid advances. Most of what is currently being called ‘AI' can be more properly labeled as machine learning. Even that term is too grandiose - but by any name, machine learning is a bona fide breakthrough in software design.
The principle is beautifully elegant. Set up a software system that answers a particular type of question with near-random results. Feed in millions of example questions, and compare the system's answers with the correct solutions. Then use these scores to refine the system's guessing process.
IBM's Candide translation system pioneered this approach in the early 1990s. It was fed millions of equivalent English and French sentences - drawn, oddly enough, from bilingual Hansard proceedings of the Canadian Parliament. Eventually, the system became adept at translating between the two languages.
Today, Google Translate works similarly. These systems don't ‘understand' language. Instead, they use ever-more-reliable statistical math to choose equivalent words and phrases. (There's also a great deal of work done around the edges, to deal with special cases, and make the system more efficient.)
Neural networks are a particularly interesting way of building machine learning systems. First, you create a tiny software ‘processing unit' that uses a simple rule to combine multiple inputs into a single output. Then, combine many of these units into a system that can take a more-complex input and create an overall result. As with other machine learning systems, sample data is fed in and the results are used by the network to adjust itself.
Neural networks allow certain types of processing to be represented elegantly. They're effective in analyzing ‘fuzzy' data, such as rough images, market statistics, or even engine noise. Google has a great example at quickdraw.withgoogle.com - a neural network that learns to recognize what's depicted by users' doodles.
Neural networks are an interesting way of building machine learning systems: Google Quick Draw, which analyzes a user's doodle drawings and attempts to correctly identify what they are, is a simple but great example.
At present, neural nets need a lot of computer power, so you're most likely to encounter them at the other end of a cloud connection, like Google's Quick Draw. The myriad of processing units of a neural network don't map well onto today's hardware, with its limited number of processor cores.
Nonetheless, neural nets have been a huge advance in tackling certain types of problems. This, and the fact that they seem to resemble the structure of the human brain, leads enthusiasts like Kurzweil to conclude that they presage a quick development of strong AI. There are problems with this view.
For a start, the human brain has close to 100 billion neurons operating simultaneously. Each neuron may connect to 10,000 others, for a total of several hundred trillion synapses. Neural networks approaching that scale have been simulated in the lab, but only by running them hundreds or thousands of times slower than real time. They've shown some interesting brain-like behaviours, but that's about it.
Still, learning systems in general are an exciting development. But before considering their impact, we have to deal with the odd fact that machine learning is not only being misleadingly called AI, it's become virtually synonymous with AI.
It is in fact quite probable that machine learning does take us closer to the development of genuine artificial intelligence. But machine learning in itself is not remotely close to what most people would think of as ‘AI.' While speaking at a Microsoft-sponsored AI panel discussion in Seoul, South Korea, Seung-won Hwang, Professor at Yonsei University, referred to machine learning as just "another paradigm of programming."
At the same AI forum, Jin Hyung Kim theorizes that "strong AI" should be possible, but that we're nowhere close to it today. "It's a long, long, long, long way to go," he predicts.
The Applications
Over-ambitious use of the term ‘AI' is particularly unfortunate in that it tends to obscure the true value of machine learning technology. These new systems are so useful exactly because they're good at something humans do badly: chewing through huge masses of data and extracting useful patterns or correlations.
This is sometimes referred to as data mining, although even that term doesn't quite suggest the full range of possibilities.
On a large scale, a security agency might collect massive amounts of user data and ‘mine it' for certain patterns of phone calls, or certain combinations of social media connections. But learning systems can also offer more personal service, when incorporated into search engines, or personal assistants like Siri and Cortana.
Perhaps the most mundane example of machine learning is to be found in e-mail software, such as Mozilla's Thunderbird. Bayesian filtering uses statistical analysis to detect spam messages with amazing accuracy, while leaving legitimate mail alone.
IBM is now selling a family of ‘Watson' products for all sorts of applications. Here are a few examples. Above is IBM Research Manager Abe Ittycheriah demonstrating how IBM Watson parses a sentence in English and "tokenizes" it to learn Korean... (Credit: IBM)
...here, VA oncologists Michael Kelley (left) and Neil Spector review a Watson for Genomics DNA analysis report, to help doctors expand and scale access to precision medicine for American veterans with cancer... (Photo Credit: Martha Hoelzer)[watsonabeibm-1.jpg]
...The Macy's On Call mobile web tool powered by IBM Watson and Satisfi allows customers to input questions in natural language about the store's product assortment, services and facilities and receive customized responses... (Credit: IBM);
...Local Motors CEO and co-founder John B. Rogers, Jr. with Olli, the first self-driving vehicle to integrate the advanced cognitive computing capabilities of IBM Watson. (Rich Riggins/Feature Photo Service for IBM)
In Thunderbird, you click a little flame icon to mark or un-mark messages as spam. The filter ‘learns' quickly from your selections. There's no need to blacklist or whitelist specific mail sources, or to create narrow rules (such as looking for the word "enlargement"). All those nuances are taken into account by the stats. Many other e-mail systems use the same approach.
Nobody would claim that the tiny bit of Bayesian software is ‘intelligent.' But it is incredibly efficient at its designated task.
Advanced artificial intelligence is often ascribed to computer games. In reality, computer-controlled opponents in games tend to operate based on relatively simple rules, hand-scripted by the game's developers. If the game is too easy to beat, these scripts may ‘cheat' by accessing information withheld from the human player. Often, the game's scoring may simply be slanted in favour of the computer.
Computer systems that play games have also not been very smart. In 1997, Deep Blue (barely) beat chess grandmaster Garry Kasparov - purely by extrapolating every possible move and counter-move, evaluating up to 330 million positions per second. Humans don't play that way. They abstract the problem, looking for patterns.
However, there have been advances. Recently, Google DeepMind's AlphaGo beat human Go master Lee Se-dol, four games to one. Go is too complex for the brute force approach, so AlphaGo combined conventional software logic with machine learning. Presumably, it digested many typical game configurations, until it was able to produce desirable moves.
It would be a stretch to suggest that AlphaGo ‘understands' Go. But the software does demonstrate the power of machine learning in attacking problems that involve massive numbers of potential data points.
Another new game-playing system, IBM's Watson, is even more interesting. Watson won a Jeopardy! match against two human champions by correctly interpreting the game's cryptic English-language clues.
Watson used multiple techniques, ranging from pre-coded logic rules to full machine learning. It came up with multiple possible meanings for each question, then chose one according to a probability score. Retrieving the answers was relatively easy. Watson had hundreds of millions of pages of reference material stored in memory, reputedly including the full text of Wikipedia. It also ran on powerful hardware: a cluster of 90 IBM servers, with a total of 720 processor cores.
To put this in perspective, compare it to the ‘hosts' of HBO's Westworld, which not only understand spoken language, but ad lib appropriate responses, while running on a computer that can fit into a human-sized skull. Obviously, we are multiple generations of hardware and software breakthroughs away from that kind of AI.
Nonetheless, Watson is an impressive achievement, and a very useful advance in natural-language processing. IBM is now selling a whole family of ‘Watson' products for all sorts of applications. Consumers may already be encountering Watson in the cloud, or in automated phone systems.
There are many practical applications for learning systems, from finance, to weather forecasting, to surveillance and security. Applications in medicine are particularly advanced, helping both with day-to-day treatment and longer-term goals.
Early attempts to use computers in medicine were focused on expert systems: essentially electronic textbooks that tried to offer an intuitive means of retrieving vital knowledge. Various other systems have been in wide use since the 1980s, able to compile lists of symptoms and match them against possible diagnoses.
Today, machine learning (including IBM's Watson systems) is finding a natural fit. Feed in our huge existing library of symptoms and diagnoses, and the system can start to recognize specific conditions. It may even be better able to deal with vague or contradictory symptoms than a human. Of course, it's unlikely that anyone today would want to be diagnosed solely by a computer. But machine learning systems can give doctors a better basis for their final conclusion.
There are lots of other possibilities. For example, learning systems can take data on weak drug effects and suggest chemical attributes that might strengthen those effects. This could help lead researchers to entire new families of useful compounds.
A very current example of machine ‘intelligence' is the quest to create a self-driving car. These projects use every trick in the book, and still have a rough road ahead.
Ford is one of many vehicle manufacturers working on autonomous cars that use machine-learning techniques and map data to enable self-driving. Seen here is the Fusion Hybrid autonomous vehicle, which has been tested in Michigan, Arizona, and California. Ford aims to launch a fully autonomous vehicle for ride-sharing by 2021.
A big enabler is radar sensing, which allows a car to ‘see' its complete surroundings in real time. This image data needs to be combined with GPS data, standard map data and detailed data compiled on specific streets. So far, self-driving cars generally operate only on streets that have been mapped to high accuracy, and extensively test-driven by humans.
Machine learning techniques are used to analyze all this data, both ahead of time and while driving, with the heaviest work being done in the cloud. Increasing amounts of data make the system more reliable, but even so, autonomous cars have trouble dealing with surprises, or edge cases. Computers still find it difficult to extrapolate from ‘experience,' to deal with radically unexpected challenges
Even so, self-driving cars are probably only a few years off. They'll rely heavily on machine learning, but they'll succeed mainly because clever humans have anticipated every possible eventuality.
The Limitations
The dark side of machine learning is the ability to take masses of statistics about our daily lives and extract more information than we'd care to have known. There's no easy way to prevent this kind of analysis, so it will become increasingly important to place tight controls on the collection of even the most innocuous personal details.
Police forces are already using ‘data-driven policing' to predict crime patterns. One clearly beneficial application is in analysis of traffic accident data, to remedy dangers that might not have been obvious to human examination. For retailers, data can help reduce the incidence of shoplifting, by identifying high-risk store areas.
On the other hand, it's a very short step from such benign uses to sweeping criminal profiling that could stigmatize wide swaths of the population. Nobody wants to find himself arrested (or assassinated) for a crime he hasn't (yet) committed, as in the film Minority Report, or the TV series Person of Interest.
In Sci-Fi, it's often the systems themselves that become dangerous. But in reality, it's humans who will decide how they're used.
Often considered "AI," voice assistants like Microsoft's Cortana are more accurately a form of data mining or learning system that puts together massive amounts of data and extracts useful patterns or correlations to offer a personal service.
The technology does have its own limitations. For example, machine learning systems are obviously single-purpose. A system like AlphaGo does one thing well. Medical diagnosis would require a separate system trained for that purpose.
Is the human brain just a collection of such systems? We can't be sure. What we do know is that the most powerful learning systems today, like Watson or AlphaGo - the ones that come closest to looking like genuine artificial intelligence, if only within narrow bounds - tend to be hybrids, in which machine learning is supported by the best procedural logic humans can devise.
Turing said that the question of whether machines can think is too meaningless to deserve discussion. Machine learning seems to reinforce this view, solving difficult problems without obviously mimicking human thought processes.
Of course, in a broader sense, all computing does mimic human intelligence in some way. But we shouldn't let our terminology jump the gun. Nobody called the first light-emitting diode a ‘high-definition display.' And nobody referred to the first TVs in the 1940s as ‘virtual reality.'
What's more, nobody in the 1950s would have predicted that HD displays and VR systems were mere decades away - while AI remained a distant goal. Today, the future of digital technology is only becoming more difficult to predict. In some areas, such as heat dissipation inside silicon chips, we're coming up against hard limits. And yet progress continues in other ways.
Like other technologies, AI will probably continue to evolve inexorably along multiple paths, each of which will spawn useful applications along the way. Ultimately, our biggest challenge isn't to make computers smarter, but to become smart enough ourselves to use them wisely.
Photo at top: One of the most recent depictions of artificial intelligence in mainstream entertainment is in the popular new HBO sci-fi western Westworld: in the series, synthetic androids known as "hosts" are able to understand spoken language and ad-lib appropriate responses, all while running on a computer that can fit into a human-sized skull. Photo courtesy of HBO Canada