It’s easy to make the case that Google (GOOGL) is at the forefront of artificial intelligence research. Their Neural Turing Machine can essentially program itself, and they’ve acquired DeepMind Technologies, a lab that has repeatedly made headlines for their AI advancements. One of DeepMind’s computers even taught itself to completely dominate classic Atari games.
We wanted to find out what the company is up to now and how their current research will play into the technology we use every day, specifically Google’s own services. We chatted with Google’s Jason Freidenfelds about the current state of natural language understanding, search and more.
When you talk to Siri or any other speech-to-text feature of your phone, all it’s doing is transcribing what you say. Google has developed and officially launched a new system that uses a muilt-layered deep learning neural network to cut down on errors by 25 percent. The whole area has still seen a hand-engineered approach to structure that looks for nouns, verbs, something to indicate it’s a question, etc., but Google is still working on developing a more sophisticated approach to natural language.
Natural language and search
Language was a big topic because it plays such an important role in so many facets of AI research. It’s interesting, Mr. Freidenfelds pointed out, that teaching machines to understand natural language has been one of the biggest challenges facing researchers thus far because, in some sense, language feels like it was the first thing that technology can handle and understand. For example, if you think about a Google search, you type words or phrases and get back relevant results. It seems simple. However, machines have always just matched keywords; when you type, “how to change a tire,” it looks for results with that phrase or synonyms like “repair.” Researchers are now trying to advance machines to actually understand natural human language instead of approach it like “a bag of words.” This will help machines generate answers to more complex questions like, “What’s a school near me that would be good for my daughter with special needs?”
This AI computer outperformed humans on an IQ test
Mr. Freidenfelds explained the neural network Google is using to build up representations of words and word sequences. With word vectors, they plot words in a space that visually depicts how words are related to each other.
“If you take all the text from very large quotes and see how they’re all related to each other in the way they’re used in real sentences, you’ll see the relationship between “Paris” and “France” is parallel to the line between “Rome” and “Italy”,” he said, emphasizing that they’re literally visually parallel.
Basically, the neural network is building up this space of words with information on how they’re used in the real world, what they mean and how frequently they appear together. It will eliminate the need for machines to manually specify knowledge—For example, come across “Paris” and understand, “it’s a city, a capital, and a capital of a country” and then do that same process for “Rome.”
Sentence shape and translation
The same can be done with entire sequences using thought vectors. A sentence has a very complex path through words, and the shape of the path represents a very abstract view of that whole.
Every single sentence has a completely unique shape, and similar sentences have similar shapes. For example, the shapes of the following two sentences would be extremely close:
I’d like to change my tire.
I want to repair my tire.
Identical sentences expressed in different languages have identical shapes. So, once a machine knows the shape of a sentence in one language, it can use that to look for the same shape in any other to get the translation.
Currently in the research phase is Google’s new system for automatically creating captions for images.
They’ve trained a neural network to recognize images very well by feeding it good examples. The image is the input and the caption is the output, and the more images they feed it, the better the captions become.
“It’s learned to create good captions for images it’s never seen before,” Mr. Freidenfelds said.
This will make images easier to organize and find, essentially creating a more sophisticated network of visuals online.
Where Google’s machine learning technology is already in use
AI has help some really cool (and some really frivolous) things happen. Google photos uses the latest in neural networks in for image search.
Their speech recognition software is using deep learning networks as well. As mentioned, it hasn’t completely overhauled the system, but rather improved it.