Artificial intelligence has been misunderstood for as long as it has existed. Herewith, the definitive Popular Mechanics guide to what A.I. is, what it isn’t, who’s responsible—and how you can get involved.
The Acoustic Model: Sounds Into Data
Your voice is measured by frequency, the wavelengths of sound at a specific moment. When you speak to Alexa, the software breaks down your command into 25-millisecond slivers, then converts each wavelength measurement into digestible numbers. The software compares those sonic signatures to its catalog of sounds until its confidence scores are high enough that it can assume that you said, “Order more dog food.”
The Language Model: Getting Meaning
Watch your phone screen while using Siri or Google Assistant, and you’ll see the transcription swap out words as you speak. That’s the software comparing the words it thinks you’ve said to its stores of example sentences, which inform how it understands syntax and vocabulary. “The language model is trained on billions and billions of words of text,” says Rohit Prasad, the head scientist behind Amazon’s Alexa division. “The web, catalogs, our assets, all of that goes into the language model that judges how likely one word is to follow in a sequence.” His example: “Play music by Sting” is far more likely than “Play Sting music by.” Also, the software considers context to be more accurate. Ask “Who is in the cast of Dark Knight?” and it will read the result for the massively popular The Dark Knight Batman movie, rather than the lesser known 2017 art film Dark Night.
Acoustic and language models constantly adjust to how people use them. That’s where A.I., specifically machine learning, comes in. “When you barge in on Alexa,” Prasad says, referring to when users have to rephrase or specify a command or ask Alexa to stop, “we know we must’ve done something wrong. That’s a cue to learn.” For example, if you say, “Play The Martian,” the device will consider whether you want the audiobook or the Matt Damon movie. If the device needs to ask which one to play, and the number of users requesting the movie far outweighs the audiobook, the device might later default to playing the movie.
What’s Coming Next
“Somebody who worked for one of these big companies told me that one of the most fascinating things you can do with all this data is to find out what people are asking about that the system doesn’t do,” says professor Alan Black of Carnegie Mellon University’s Language Technologies Institute. “When the thing says, ‘I can’t do that,’ you find the words recognized, and discover that it’s something really interesting.” Responding to unpredictable requests is part of the goal for student researchers competing for the $2.5 million Alexa Prize. The challenge: produce a chat bot that will converse with a human—ask intelligent follow-up questions, add new information—for 20 minutes. That’s not difficult for a customer-service chat bot that’s only concerned with khakis, but in this situation, as with conversations between humans, there are digressions and spontaneity. A program capable of that will mean a huge leap for A.I.
Source : Popular Mechanics