(Bloomberg) -- Real-world artificial-intelligence applications are popping up in unexpected places—and much sooner than you might think.
While winning a game of Go might be impressive, machine intelligence is also evolving to the point where it can be used by more people to do more things. That's how four engineers with almost zero knowledge of Japanese were able to create software, in just a few months, that can decipher handwriting in the language.
The programmers at Reactive Inc. came up with an application that recognizes scrawled-out Japanese with 98.66 percent accuracy. The 18-month-old startup in Tokyo is part of a growing global community of coders and investors who are harnessing the power of neural networks to put AI to far more practical purposes than answering trivia or winning board games.
"Just a few years ago, you had to be a genius to do this," said David Malkin, who has a Ph.D. in machine learning but can barely string two Japanese sentences together. "Now you can be a reasonably smart guy and make useful stuff. Going forward, it will be more about using imagination to apply this to real business situations."
Artificial intelligence was once the exclusive playground of Google Inc., Facebook Inc. and other tech leaders. Now, any deep-learning startup can access cloud-based platforms, with Microsoft Corp., Nvidia Corp. and Amazon.com Inc. selling AI like a utility.
Reactive’s technology shows how even small teams can devise complex applications with little expertise in a given field. The hard part may be figuring out how to make money. To that end, Reactive intends to help Japanese schools grade papers—a prosaic exercise that may change the game in a country where tests are still handwritten.
Malkin and his colleagues, Joe Bullard, Philippe Remy and Philip Irri, who between them have two Master's degrees and a PhD, are making rapid progress. Bullard showed the group's program to AI enthusiasts during a social gathering at Google’s Japan headquarters earlier this year, and it performed impeccably—until his own linguistic limitations got in the way.
"Just so you know I’m not cheating, what’s your favorite character?" he asked the room. Someone called out "ba," a phonetic character. Knowing laughter erupted as Bullard struggled to write the symbol.
While handwriting recognition might be considered deep-learning 101, Japanese is a whole other ballgame. That's because the language includes symbolic characters such as kanji, which is composed of elements that can be read independently, making it difficult to know where one ends and another begins. There are also more than 2,000 common pictograms made up of dozens of strokes. The trick is to tackle one squiggle at a time. Reactive’s algorithm queries the neural network for a match, adds another stroke and repeats, all the while refining the probability of an accurate hit. The startup trained its model on about 1.8 million characters.
"The fact that this technology can side-step domain expertise gives it a big advantage in speed and scalability when it comes to business applications," said Seishi Okamoto, a project director at Fujitsu Laboratories Ltd., which is developing software to read Chinese. "Deep learning for handwritten Chinese character recognition is already catching up to human capabilities and will probably eclipse them." While Reactive's technology has been shown publicly at events such as that held at Google, the data hasn't been independently verified or peer-reviewed.
Unlike a typical program built around rigid rules, deep-learning AI is modeled on how humans process information. Given enough data as inputs and a set of desired outputs, neural networks figure out what goes in the middle. This allows them to find solutions that have bedeviled traditional approaches, like interpreting speech or tagging images.
"There are several factors at play: high-performance computing becoming essentially a commodity, availability of massive datasets and improvements in basic science," said Yoshua Bengio, a computer-science professor at the University of Montreal who co-authored some of the field’s founding papers. "Democratization of tools is also making it easier for second-level users to develop new applications and products."
All of this is likely to speed up the evolution of artificial-intelligence technology. Investments in AI startups tackling everything from education to retail and agriculture reached $310 million in 2015, almost a seven-fold increase in five years, according to research firm CB Insights.
And once built, a neural network doesn’t have to be limited to language applications. In their spare time, the four Reactive engineers showed the program 5,000 dresses downloaded from Google Images, then gave it a picture of a woman in a revealing outfit. "Sexy clothes," the software responded.
