What Is Google DeepMind?

How deep learning is embedded in the products you use

google deepmind
DeepMind uses neural networks and deep learning to train artificial intelligences.

Yuichiro Chino / Getty Images

DeepMind may refer to two things: the technology behind Google’s artificial intelligence (AI) project, and the company that is responsible for it. The company DeepMind is a subsidiary of Alphabet, the parent company of Google.

DeepMind's artificial intelligence technology has found its way into a variety of Google projects and devices. If you use Google Home or Google Assistant, then your life has already intersected with DeepMind in some way.

How and Why Did Google Acquire DeepMind?

DeepMind was founded in 2011 with the goal of “solving intelligence, and then using that to solve everything else.” The founders tackled the problem of machine learning using insights from the field of neuroscience. Their goal was to create powerful general-purpose algorithms that would be able to learn and reprogram themselves, rather than needing to be manually programmed by humans.

Several players in the field of AI were impressed by the talent of the DeepMind team. In 2012, Facebook made a play to acquire the company. That deal fell apart, but Google swooped in and acquired DeepMind in 2014 for about $500 million. DeepMind then became a subsidiary of Alphabet during the Google corporate restructuring of 2015.

Google’s main reason for buying DeepMind was to jump-start its artificial intelligence program. While DeepMind’s main campus remained in London, England, an applied team was dispatched to Google’s headquarters in Mountain View, California. That team was to work on integrating DeepMind AI with Google products.

What Is Google Doing With DeepMind?

DeepMind’s goal of solving intelligence did not change when it handed the keys over to Google. Work continued on deep learning, which is a type of general machine learning program. This compares with earlier AIs like the Deep Blue computer, which in 1996 famously defeated chess Grandmaster Gary Kasparov. Such computers excelled at domain-specific tasks but were minimally useful outside of those domains. DeepMind, on the other hand, was designed to learn from experience.

DeepMind’s artificial intelligence has learned how to play video games like Breakout better than the best human players. In 2016, a DeepMind-powered program called AlphaGo defeated a world champion Go player—a milestone for the fact that Go is considerably more complicated than chess. In addition to pure research, Google has integrated DeepMind AI into its flagship search and mobile devices, including Google Home and Android.

How Does Google DeepMind Affect Your Daily Life?

DeepMind’s deep learning tools have been implemented across the entire spectrum of Google products and services. If you use Google, there’s a good chance you have interacted with DeepMind in some way.

Some of the most prominent uses for DeepMind AI include speech recognition, image recognition, fraud detection, spam identification, handwriting recognition, translation, Google Maps Street View, and Local Search.

Google’s Super-Accurate Speech Recognition

Speech recognition, or the ability of a computer to interpret spoken commands, has been around for a long time. Virtual assistants like Siri, Cortana, Alexa, and Google Assistant have brought the functionality closer to our daily lives.

In the case of Google’s voice recognition technology, deep learning has been deployed to great effect. Machine-learning has allowed Google’s voice recognition tech to achieve an impressive level of accuracy for the English language, to the point where it is as accurate as a human listener.

If you have any Google devices, like an Android Phone or Google Home, this has a direct impact on your life. Every time you say, “Okay, Google” followed by a question, DeepMind flexes its muscles to help Google Assistant understand what you are saying. Unlike Amazon’s Alexa, which uses eight microphones to understand voice commands, Google Home’s DeepMind-powered voice recognition system only requires two.

Google Home and Assistant Voice Generation

Traditional speech synthesis uses something called concatenative text-to-speech (TTS). When you interact with a device that uses this method of speech synthesis, it consults a database full of speech fragments and assembles them into words and sentences. This results in a strangely inflected speech pattern, and it is usually clear that the speaker is not human.

DeepMind tackled voice generation with a project called WaveNet, which was meant to make artificially-generated voices sound more natural. WaveNet relies on samples of real human speech but does not use the samples to synthesize new voices. Instead, it analyzes the samples of human speech to learn how the raw audio waveforms work. This allows the program to speak different languages, use accents, or even be trained to sound like a specific person. Unlike other TTS systems, WaveNet generates non-speech sounds like breathing and lip-smacking to render an even more realistic vocal profile.

If you want to hear the difference between a voice generated through concatenative text-to-speech, and one generated by WaveNet, DeepMind has some interesting voice samples that you can listen to.

Deep Learning and Google Image Search

Without artificial intelligence, searching for images relies on context clues like tags, nearby text, and file names. With DeepMind’s deep learning tools, Google Image Search was able to learn what various people and objects look like, allowing you to search your own images and get relevant results without needing to tag anything.

For example, if you search for “dog,” Google will pull up images of your dog that you took, even if you never labeled them. This is because it was able to learn what dogs look like, in much the same way that humans learn what things look like. And, unlike Google’s dog-obsessed Deep Dream, it’s more than 90 percent accurate at identifying all sorts of different images.

DeepMind in Google Lens and Visual Search

One of the most stunning developments of DeepMind is Google Lens, a visual search engine that allows you to take a picture of an object in the real world and instantly pull up information about it.

While the implementation is different, this is similar to the way that deep learning is used in Google Image search. When you take a picture, Google Lens is able to look at it and figure out what it is. Based on that data, it can then perform a variety of more advanced actions.

For example, if you take a picture of a famous landmark, it will supply you with information about the landmark. If you take a picture of a local store, it can pull up information about that store. If the picture includes a phone number, Google Lens is able to recognize the information and give you the option of calling the number.