Google Announces an AI That Surpasses Humans at Programming AIs

Thursday, 25 May 2017 - 12:44PM
Technology
Artificial Intelligence
Thursday, 25 May 2017 - 12:44PM
Google Announces an AI That Surpasses Humans at Programming AIs
Image credit: Google Developers
It's telling that the phrases 'neural network' and 'deep learning' have become just another pair of tech buzzwords, despite the fact that what they're really talking about is creating brains. Fans of Isaac Asimov's robot stories will be familiar with his famous 'positronic brain', which were the basis for robots who could think and learn like humans. Even when they inevitably learned to bypass the Three Laws, though, Asimov's robots weren't a huge threat. Now, however, scientists and developers have begun to create neural networks that can create even better neural networks, ad infinitum—and they've already begun to surpass humans at doing it.

We're just gonna leave this here:


At the recent I/O 2017 Event, Google CEO Sundar Pichai announced that, among other advances, Google's machine learning tech had recently surpassed humans in the field of image recognition (Skip ahead to 10:10):



This is a cool announcement, but it's only the tip of the iceberg—according to Google's Research Blog, their new technique for designing neural networks and using deep learning, AutoML, can create more efficient and powerful neural networks on its own, which are actually better than those designed by humans:

Opening quote
We've applied [AutoML] to two heavily benchmarked datasets in deep learning: image recognition with CIFAR-10 and language modeling with Penn Treebank. On both datasets, our approach can design models that achieve accuracies on par with state-of-art models designed by machine learning experts (including some on our own team!).
Closing quote


How does AutoML work? Essentially, one neural network is made a 'parent' network, which proposes various 'child' networks to try to meet a goal, like accurately predicting what word will come next in a sentence. Based on feedback from humans, the 'parent' network acts as an intermediary to design new networks that fulfill the guidelines and goals they're given. This is a huge deal, because creating a neural network (which is basically modeled after the structure of the human brain, complete with layers of neurons) from scratch is an incredibly difficult and time-consuming task—according to the Blog, a typical 10-layer neural network produces about 10 billion candidate networks to sort through and assess.



So what uses does AutoML have, besides being the perfect way to produce an apocalyptic sci-fi scenario we've been warned of time and time again? The answer is actually pretty exciting: with AutoML acting as an intermediary, non-experts in machine learning and neural networks could conceivably start to create custom AIs for all kinds of different tasks, from learning what books to recommend to how to register human facial expressions and act accordingly. The most striking piece of knowledge, however, is that this kind of 'deep learning' is predicted to be infinitely scalable—meaning that the more data you feed it and the more processing power you give it, the better it's going to get.

We've already lived to see the day when automated programs unseat humans at jobs we thought required a human touch—from travel agent to taxi driver. Now we may see AI, in the form of neural networks, creep ever deeper into everyday life, constantly improving until it becomes the bedrock for business, trade, entertainment, travel, and how we interact with the world. It could even take over writing articles about itself. 

I, F0R 0NE, WELC0ME 0UR R0B0T 0VERL0RDS.
Science
Science News
Technology
Artificial Intelligence
No