Google’s Neural Network at Work

On June 26, 2012 Google appeared at the International Conference on Machine Learning and demonstrated a neural network it built. A quick recap on that, the network is meant to create a functional artificial intelligence, and it works. They showed the computer over 10 million different YouTube video thumbnails over three days and gave it a list of 20,000 different items. The computer, despite none of the 20,000 items being cats or anything about them, started learning how to identify a cat and what characteristics it had using a deep learning algorithm. The news made headlines across the tech industry, a computer had learned, with no human intervention, what a cat was and how they work. Fast forward a few months to today.

Google didn’t just stop with the cat stuff; they are pushing this budding neural network further now to enhance their own products. They have already had the system examine their voice recognition system (Google Now is what it’s called, it is Google’s answer to Siri) and the neural network was able to improve the voice recognition accuracy and efficiency by 25%. Instead of merely listening to what the person is saying and guessing the nearest word, it examines the context of the sentence as well as the word that it wasn’t sure of and sorts out the most likely thing you were talking about. So let’s say you’re in southern California and you need directions to Modesto but some passing car interrupts the phone from hearing you say Modesto and it is now faced with wondering if you said Modesto or maybe Monaco (in the French Riviera). Based on previous inputs and your location, the system will go with Modesto as it would make more sense. The system learns and is still learning.

    Google has been applying it to some of the Google Maps cars that drive around snapping street level pictures of everything. The images that come back are usually analyzed by humans for house numbers and such, but the neural network is able to identify those things with a much higher accuracy rating which will improve location accuracy on the maps. One of the goals for the project is for the system to reach animal or even possibly human levels of intelligence, but it’s definitely not there yet. The neural network is small, and despite the victories it has had, it is still a long road before we see Skynet type reasoning. For immediate use though, the Google self-driving cars and Google Glasses seem to be ideal areas to put this system to work. The self-driving cars have been approved for use by California and a system like this could very easily help improve the ability for the cars to understand their surroundings and even maintain a better working car based on engine sensors. The future is coming and it looks like neural networks are only getting smarter from here on out.

[Technology Review, Wired]