GOOG Alphabet Inc

By Daniela Hernandez 

Artificial-intelligence systems can do increasingly complex tasks but they can't yet figure much out on their own without help from humans.

In a paper published Wednesday in the journal Nature, researchers at Alphabet Inc.'s Google DeepMind describe experimental software that they say gets closer to that goal and could be more accurate and less costly than current systems.

Their prototype learned to navigate the London Underground and plan the shortest route to a destination. It could also figure out how imaginary people were related. For both challenges, it drew on and retained knowledge it had learned previously, suggesting that the new technology could learn from experience, according to Alex Graves, a DeepMind research scientist and one of the study's lead authors.

"There's a lot of things it could be used for," he said. One obvious future application is "chatbots," software that answers questions autonomously, he said.

The new DeepMind prototype couples so-called artificial neural networks -- which are widely used for image and speech recognition -- with an external memory. Data can be read from and written to memory, like in a traditional computer. Dr. Graves likened it to a scratchpad a person could write on, reference and update as he or she made decisions and learned new facts.

The researchers trained their software, dubbed a differentiable neural computer, to answer questions, solve a simple puzzle, and optimize subway rides. Its error rate in the question-answer test was 3.8%, according to the paper.

The new DeepMind prototype outperformed more traditional neural networks, but required half as many examples to learn, according to the study. Data is among the most expensive ingredients needed to train algorithms.

Current artificial neural networks can retain only small amounts of information as they scan data. This limits their ability to perform tasks that require long-term planning or historical knowledge, like navigating new environments without a human teacher or engaging in nuanced conversations peppered with jokes and cultural references.

"Right now, that's beyond reach," said Bruno Olshausen, the director of the University of California, Berkeley's Redwood Center for Theoretical Neuroscience. "With memory, you're giving it a leg up. ... It's still a long ways off. [But] you're one step closer to a truly autonomous system." Dr. Olshausen wasn't involved with the new work.

Over the past two years, the field of memory-augmented neural networks has blossomed, according to artificial-intelligence experts.

In late 2014, DeepMind unveiled so-called Neural Turing Machines, a similar, but less flexible, technology than the new prototype system. Soon after, Facebook Inc. researchers published a paper describing Memory Networks, a system also coupled to an external memory. A spokesman for Facebook declined to comment.

These evolved from a special breed of software called long short-term memory networks that have a short-lived built-in memory. Such systems already power real-world applications like voice search and automated transcription.

But systems with external memories like DeepMind's neural computer haven't yet busted out of the lab, and much work likely is still necessary to make them commercially relevant, said Yoshua Bengio, the director of the University of Montreal's Montreal Institute for Learning Algorithms, who does research in this area but wasn't involved in developing the new neural computer.

The DeepMind team doesn't yet know what steps they'll take next to further develop the technology, said Dr. Graves. "We're doing academic research here," he said. "But we're confident it will be useful in the future."

Write to Daniela Hernandez at daniela.hernandez@wsj.com

(END) Dow Jones Newswires

October 12, 2016 13:18 ET (17:18 GMT)

Alphabet (NASDAQ:GOOG)
Historical Stock Chart
From Mar 2024 to Apr 2024

Alphabet (NASDAQ:GOOG)
Historical Stock Chart
From Apr 2023 to Apr 2024