Google is preparing to unleash a wave of apps that get intelligence from its mammoth machine learning models.
The apps will all rely on the neural networks Google has been developing internally to allow its systems to automatically classify information that has traditionally been tough for computers to parse. This includes human speech or unlabeled images, said Jeffrey Dean a fellow in Google's Systems Infrastructure Group who helped create MapReduce and GFS, to the GigaOm Structure in San Francisco on Wednesday.
"I've been working on a machine learning system for the last couple of years that is using biologically inspired neural networks," Dean said. "These kinds of models are very useful in a whole bunch of different domains."
Machine learning uses neural networks that evolve through hierarchies of successively more specific stages to gain sensitivities for particular characteristics of data. One Google project from mid-2012 that used the tech giant's DistBelief machine learning tech, proved that the internet really is made of cats, when part of the Chocolate Factory's neural network developed an appreciation of felines after being fed a diet of YouTube thumnbails.
Now, Google is planning many more applications that make use of the technology. "We deployed a speech detector on Android that drops our error rate by a significant amount and a lot of that is attributable [to machine learning]," Dean said. "A lot of apps we haven't deployed yet that are trying to use language understanding for these kinds of models."
The wide rollout of this technology will have major ramifications for consumers of Google's services, Dean said, and could become a dominant approach for cracking certain classes of problems.
"I think this kind of perceptual machine learning is going to significantly change how people interact with devices," he said. "Speech recognition is now reliable enough that you can build complicated [features] around just speech."
Machine learning is so important to Google that it is one of the areas that the company's dedicated research wing works on. Many within the company foresee a combination of complex finely-tuned neural networks and vast quantities of user data as being one of the best ways to create and train weak artificial intelligences.
But the technology requires advanced hardware, Dean said, noting that the models are "intensive for floating point operations" when training. If you're Google, that means you need to spread the computation across thousands of CPU cores, and if you're other, smaller companies, it requires low-level coding to take advantage of GPUs, as a gang of Stanford Academics have recently done.