The world as you know it is already run by A.I.
The world as you know it is already run by A.I.
When people think of artificial intelligence (A.I.), we immediately jump to any one of the popular Hollywood tropes of robots taking over the world and enslaving humankind after attaining self-awareness. Greg Corrado, senior research scientist and co-founder of Google’s deep learning team thinks that this scenario is still just fantasy and should remain in Hollywood.
Why should we listen to Greg? Well, his title isn’t just for show. Outside of Google, his research work borders between artificial intelligence, neuroscience and scalable machine learning, he’s published papers on the topics. He’s also worked in IBM Research on neuromorphic silicon devices and large scale neural simulations, a topic that we’ve also written about before.
(see HWM October 2015, Emulating the Human Brain). So, if Skynet or the Matrix is pure fiction, what is A.I. realistically? What if I were to say that you’re already using and benefiting from A.I., or to be bold, relying on it for your everyday tasks?The Machine Learning RevolutionDevelopers today prefer the term machine learning, and it’s not just nomenclature. According to Google, machine learning is a revolution that’s changing the way software is developed.
In traditional programming, software is bound by explicit rules. For instance, as little as five years ago, when your mail client junks an email into the spam folder, it probably did so because the email contained pre-determined keywords – such as ‘viagra’. And it would continue to do so even though you may have a legitimate prescription for said drug.
You would then have to specify another explicit rule to exclude emails from your doctor, for instance. Machine learning on the other hand is a development model that challenges programmers to write programs that learn by example; similar to how humans learn from trial and error. So, while machine learning apps have a higher rate of error at the beginning, given enough time andChansample data, it can only become more accurate.
Google is of course one of the biggest proponents of machine learning and all of its products have been designed with this development model in mind. Gmail’s spam filter for example, is so accurate today that it intercepts 99.9% of all spam (according to a Google Gmail Blog, July 2015). And every time you receive mail, mark or unmark spam, Gmail continues learning. The same goes for speech and text recognition. You may have remembered when Apple launched Siri in 2011, Siri wasn’t very good at recognizing accent.
In fact, we’d say downright terrible. It is amazing that in just four short years, Siri doesn’t just understand accents, but multiple languages. Google’s own speech recognition engine, used in products such as Now and Translate has improved transcription accuracy by 20% with Translate’s average error rates reduced from 23% to just 8% today.
Deep Learning, A.I. of todayNow, speech and text analysis is a straightforward example of how machine learning is used to continuously improve software accuracy, but that’s not exactly intelligence is it? Well, then there’s deep learning, an advanced class of the machine learning model that tries to emulate a neural network, like our brains.
This artificial neural network comprises of many ‘layers’ of trainable machine learning models that work on their own functions, but collaborates and shares information with other layers to derive answers to more complex questions. An example of a service that is designed with deep learning is Google Photos, which was launched in May this year.
Computer vision, as Google calls it, is how Google Photos can “intelligently” identify information from your photos and automatically group them, tag them and display the relevant search results. Google Photos isn’t just able to identify a dog from a cat, it is even able to tell apart a Labrador from a French bulldog.
The deep learning model used by the Google Photos team is codenamed Inception and it contains 22 different layers that look at all the different things in a photo such as lines, color and shape. Why 22? We asked Chris Perry, Google’s Photos Search and Analytics lead, and the answer is quite pedestrian. Apparently, it’s because 22 was the optimal number of layers for what they were trying to achieve.
Tomorrow could be 25, next year could be 50, there’s no hard and fast rule. Also, each Google project uses their own models, so it’s not a magic number or anything. But that’s just the start. Besides being able to identify subjects within pictures, Inception is also used to learn about relationships between subjects and comprehend surroundings.
For example, you and I can look at a picture and immediately identify it as a birthday party. A computer would only be able to “see” four people and a cake, with no further context. Inception will try to guess it is a birthday through deep learning, giving computers a more realistic sense of achieving artificial intelligence.
Other Google apps that benefit from deep learning neural networks include Translate, which combines both computer vision and text recognition to translate images to-and-from different language in real time, and Smart Reply, a new feature for Inbox that is capable of analyzingHWM31incoming email and suggesting actual context-relevant replies. Of course, Google being Google, they have the advantage of an entire internet’s worth of searches and images to crawl through to train their deep learning networks.
And in November 2015, Google released TensorFlow, an entirely open sourced machine learning library for the world to use, but they aren’t the only ones. Microsoft has Project Azure, which powers Cortana and the Oxford facial recognition engine. Earlier last year, Microsoft tried to guess your age from a picture and in November, released a new tool that tried to guess your emotions.
While these advances may bring back the fear of Hollywood’s version of A.I., we need to understand that machine learning isn’t true intelligence, just a more complex method of deriving answers. What you may want to be wary of is privacy, since the only way for machine learning systems to become better is to continuously have input, input that you choose to provide.
"The deep learning model used by the Google Photos team is codenamed Inception and it contains 22 different layers that look at all the different things in a photo such as lines, color and shape.”
