Year in artificial intelligence: Most impressive AI and machine learning accomplishments

The Most Impressive Things A.I. Accomplished This Year

The Most Impressive Things A.I. Accomplished This Year

Future Tense
The Citizen's Guide to the Future
Dec. 28 2017 9:15 AM

It Was a Big Year for A.I.

AlphaGo beat the world's best Go player in May. A few months later, a new version beat the human-defeating version 100 to 0.


2017 has been a booming year for the field of artificial intelligence. While A.I. and data-focused machine learning have been around for decades, the algorithmic technologies have made their presence known in a variety of industries and contexts this year.

Microsoft UK’s chief envisioning officer Dave Coplin has called A.I. “the most important technology that anybody on the planet is working on today,” and Silicon Valley companies seem to have taken that to heart: They’ve been hiring A.I. experts right and left, and with those in short supply, they’ve started teaching employees the fundamentals of A.I. themselves.


Not every A.I. achievement has been met with admiration and applause, though. Some are worried about the human prejudices that are being introduced into A.I. systems. ProPublica found in 2016, for example, that the software algorithms used to predict future criminals were heavily biased against black defendants. And earlier this year, Facebook came under fire for the algorithmically generated categories advertisers could use to target users, which included hateful groups and topics such as “Jew hater.” Situations like these have prompted experts to urge companies and developers to be more transparent about how their A.I. systems work. However, in many other cases—especially of late—A.I. has been used to good end: To make discoveries, to better itself, and to help us expand beyond the limits of our human brains.

A.I. Spotted An Eight-Planet Solar System

Successful astronomical discoveries often center around studying data—lots and lots of data—and that is something A.I. and machine learning are exceedingly good at handling. In fact, astronomers used artificial intelligence to sift through years of data obtained by the Kepler telescope to identify a distant eight-planet solar system earlier this month. This solar system now ties our own for the most known planets circling its star, in this case Kepler-90, located more than 2,500 light years away.

From 2009 to 2013, the Kepler telescope’s photometer snapped 10 pixel images of 200,000 different stars every half hour in search of changes in star brightness. If a star dimmed and brightened in a regular, repeating pattern, that could be an indication that it has planets orbiting. (You can also use that information to estimate the size and length of orbit of a planet circling a particular star.) University of Texas at Austin astronomer Andrew Vanderburg and Google software engineer Christopher Shallue developed the neural network that made the discovery using 15,000 known exoplanet indicators. They zeroed in on 670 stars with known exoplanets, but focused specifically on weak signals—smaller exoplanets previous researchers may have missed. The planet the duo discovered, dubbed Kepler-90i, appears to be the third planet orbiting its star, much like our own Earth.


Beat The World Champion Go Player

Google’s DeepMind researchers developed an A.I. that plays the ancient, complex Chinese strategy game of Go. The initial version defeated the world’s best Go player in May, but that wasn’t enough. A few months later, Google developed a new version of this AlphaGo A.I.: AlphaGo Zero. This A.I. achieved a superhuman-level Go-playing performance—it beat the original AlphaGo A.I. 100 to 0.*

Bested Poker Pros at No-Limit Texas Hold’Em

An A.I. developed by Carnegie Mellon’s computer science department recently beat professionals at one of the most difficult styles of poker, no-limit Texas Hold’em. Unlike strategy games like chess and Go, poker is what’s considered an “imperfect-information game” because the player must make decisions even as some information is hidden. On top of that, it’s not just making moves—it’s knowing when to bluff, too. In a 20-day competition with a $200,000 prize pool and 120,000 total poker hands played, Carnegie Mellon’s A.I., Libratus, beat the world’s top poker professionals.


And Taught Itself To Program

Artificial intelligence not only made some notable discoveries and competitive successes this year. It also excelled in a different area—making its programmers obsolete. We exaggerate: Several different artificial intelligence programs (including ones developed by Google, Microsoft, and Facebook) learned how to write basic code at a level that could help non-programmers with complicated spreadsheet calculations or reduce some of the tedium that experienced developers have to deal with.

Microsoft’s A.I., DeepCoder, might be considered the most basic of the three, although it’s still an incredibly complicated feat. This A.I. can understand a mathematical problem you need to solve, look through existing examples of code for similar problems, and then develop a code-based solution. DeepCoder could eventually be useful for those who can’t or don’t want to learn to code but need to use a code-based solution for computations (for example, tricky spreadsheet calculations). The solutions are relatively simple and, in terms of solution and structure, are based on situations the A.I. has experienced before. They usually end up being between three and six lines of code total.

Google’s program, in contrast, taught itself to program machine learning software and, in one case, learned to recognize objects in photos—a much more challenging task. Named AutoML, the program ended up achieving a 43 percent success rate at its task—4 percentage points better than the code developed by its human peers. AutoML’s biggest benefit, though, is in automating the process of developing machine learning models, a process that’s normally time consuming for human machine learning experts.

And then there’s Facebook’s self-taught chatbots, which fall on a slightly different scale of self-taught abilities. The two A.I. agents, Bob and Alice, started out speaking in English but then...developed their own language to speak in. “Agents will drift off understandable language and invent codewords for themselves,” said Dhruv Batra, visiting research scientist from Georgia Tech at Facebook A.I. Research, in an interview with FastCo Design. While this got a lot of blowback in the press (“creepy” was a common headline descriptor), it’s actually a fairly common occurrence. A.I. systems evolve using a rewards-based system, and if there’s no benefit from a particular course of action, they’ll try something else instead. Still, the Facebook researchers eventually shut down the A.I. bots since their goal was to create entities that will eventually interact with people—there was no Her-style ending for these digital acquaintances.

*Correction, Dec. 28, 2017, at 4:40 p.m.: This post originally misstated AlphaGo Zero had beaten the original AlphaGo 100 to 1. It actually beat it 100 to 0.

Future Tense is a partnership of SlateNew America, and Arizona State University.

Christina Bonnington is a technology writer whose work has appeared in Wired, Refinery29, the Daily Dot, and elsewhere.