Read more: “Instant Expert: Artificial intelligence“
It might not be obvious, but you interact with AIs every day. They route your phone calls, approve your credit card transactions, prevent fraud and automatically trade stocks in your mutual fund. What’s more, they can help your doctor interpret test results. But you won’t think of these programs as having a human-like intelligence
Spam hunters
Over 90 per cent of all email sent today is spam. If that were reflected in the contents of your inbox, email would be unusable. Your main protection against the purveyors of miracle pills and work-at-home schemes comes from spam filters that rely on machine learning. As the term implies, the spam filter learns from its environment and from how people treat their emails.
Advertisement
“90 per cent of email is spam, but you are protected from the purveyors of pills by a branch of AI called machine learning”
Individual email users provide the gold standard by correctly identifying and labelling messages in their inboxes as “spam” or “not spam”. The program uses this information to break down each message into features. A feature can be an individual word, a two-word or multiword phrase, the time of day the message was sent, or the computer that sent it. Those features can then help the program decide whether or not an incoming message is spam. For example, suppose it contains the phrases “lowest prices” and “discreet packaging”. The AI will refer to global statistics that tell it that these phrases appear in 8 per cent and 3 per cent of spam, respectively, but only in 0.1 per cent and 0.3 per cent of legitimate messages. After making some assumptions about the independence of features – and applying a formula called Bayes’ rule, which assesses the probability of one event happening based on observations of associated events – it concludes that the message is 99.9 per cent likely to be spam.
But the most important thing spam filters do is update their models over time based on experience. Every time a user corrects a mistake, perhaps by rescuing a legitimate message from the junk-mail folder – the system updates itself to reflect the new reality. So programmers do not need to specify step-by-step instructions for identifying a spam message. They need only build a general learning system and expose it to examples of spam and genuine emails. The software does the rest.
Universal translators
On any given day, Google translates more text than all the professional human translators in the world decipher in a year. Google handles 65 languages, translating in either direction between any pair – over 4000 possibilities altogether. It could not do this without a field of AI called probabilistic reasoning.
In the early days, linguists built translation systems based on bilingual dictionaries and codified grammar rules. But these fell short because such rules are inflexible. For example, adjectives come after the noun in French and before the noun in English – except when they don’t, as in the phrase “the light fantastic”.
In the last decade, translation has shifted from rules that are hand-written by human experts to probabilistic guidelines that are automatically learned from real examples.
Another key aspect of machine translation is the computer-human partnership. Modern machine translation systems start by gathering millions of documents from across the internet that have already been previously translated by humans.
While machine translation is not yet perfect, it is improving at a steady pace as accuracy increases and more languages are added. Google is working on a universal translator app called Translate. Speak into your Android phone’s microphone, and the app will read back what you’ve said, . The person you are talking to can reply in their own language.
AI hits the road
If, on your way to Las Vegas, you pass a car with red licence plates and an infinity symbol, be warned that the car is driving itself. Earlier this year, the state of Nevada issued the .
Will self-driving vehicles catch on in the rest of the world? Until now, driving has been a task best left to humans precisely because it involves so many variables: is the approaching car going at 60 or 70 kilometres per hour? Could there be another vehicle out of sight around the corner? If I attempt to pass the car ahead will its driver speed up? And many more.
For AI, the actual driving has been the easy part, at least on highways. In 1994, two driverless Mercedes-Benz cars fitted with cameras and on-board AI drove .
However, most driving takes place in cities, and that’s where it becomes tough for AI, which until recently was unable to negotiate the unwritten rules of city traffic. For example, when Google’s researchers programmed an autonomous vehicle to faithfully give way at an intersection as specified in the driver’s manual, they found that the self-driving car would often never get a chance to go. So they changed the car’s behaviour to make it inch forward after it had waited a while, signalling its intent to move ahead.
Another major source of uncertainty for a self-driving car is locating itself in space. It can’t rely solely on GPS, which can be off by several metres, so the AI compensates by simultaneously keeping track of feedback from cameras, radar, and a range-finding laser, crosschecked against GPS data. An average of these imperfect locations provides a highly accurate measurement.
But AI is not restricted to driving. In recent model cars, an AI program automatically adjusts the fuel flow to make it more efficient and the brakes to make them more effective.
Cutting-edge self-driving cars bring together many branches of AI, and people are starting to accept the idea. With special permission, fleets of self-driving Google cars have already negotiated hundreds of thousands of kilometres on California’s highways and busy city streets with no human intervention. Florida and California have now , and others may soon follow.
Data miners
In 2011, IBM introduced the world to Watson, a question-answering supercomputer with the ability to make sense of questions posed in everyday language, and answer them accurately. Watson’s 3000 networked computers were loaded with millions of documents which it could access instantly to answer almost any question.
The company set this computational brawn loose on Jeopardy!, an American quiz show famous for posing questions that are far from straightforward. The game is much more complex than chess: Jeopardy! requires not only the sum of all human knowledge, but also the ability to understand the puns and wordplay that constitute the game’s questions.
The branch of AI primarily responsible for Watson’s smarts is probabilistic reasoning: the ability to extract full understanding from a combination of incomplete information. Before the contest began, Watson was loaded with text from encyclopedias, web pages, other reference works and previous Jeopardy! games. IBM then divided the Watson program into a committee of about 100 subprograms, each in charge of its own specialised area, for example, famous authors. After the experts had scoured their own databases, Watson pooled the knowledge of all its component experts and selected the answer with the highest probability of being correct. The machine defeated two human champions.
But Jeopardy! championships are not Watson’s true calling. IBM plans to spin off the supercomputer’s game show success into more serious work, such as providing doctors, businesses, farmers, and others with time-critical information.
