DeepMind

DeepMind is a well-known AI research institution under Google and has released famous models such as AlphaGo, AlphaFold, and WaveNet. There are very eye-catching research results in fields such as deep learning and reinforcement learning. This article mainly introduces the relevant situation of DeepMind, a research institution.

DeepMind's headquarters in King's Cross, London

DeepMind company information

The main information about DeepMind is as follows:

Company type: Research company

Industry: Artificial Intelligence

Date of establishment: September 23, 2010

Founders: Demis Hassabis, Shane Legg and Mustafa Suleyman

Headquarters: London

Official website: https://www.deepmind.com/

DeepMind’s founding history

DeepMind is an AI research institution established in London in 2010 by Demis Hassabis, Shane Legg, Mustafa Suleyman and others. In 2014, Google acquired DeepMind. DeepMind released AlphaGo in 2016 and became known to the world after defeating the human Go champion Lee Sedol. DeepMind was originally named DeepMind Technologies and was renamed Deep Mind after being acquired by Google.

DeepMind has been studying teaching AI to play games since its establishment, including Breakout, Pon and Space Invaders. This is what they see as a way to create general artificial intelligence. Early investors in DeepMind include entrepreneurs such as Elon Musk (this old man is also the boss of OpenAI, so he has a vicious vision). In 2013, Facebook began to contact DeepMind in the hope of acquiring it, but in the end, in 2014, Google spent US$500 million to acquire DeepMind.

Overview of DeepMind’s research results

As a research institution under Google, DeepMind has a large amount of output in the field of artificial intelligence. Currently, DeepMind’s research areas include the following 7 areas:

Deep Learning

Control and Robotics

Reinforcement Learning

Sciences

Theory and Foundations

Unsupervised Learning and Generative Models

Neuroscience

DeepMind has a lot of very influential research. We will select some representative results for description.

AlphaGo

AlphaGo is the world's first computer program to defeat a professional human Go player. There are only two participants in Go, and each person takes turns playing chess on the chessboard. Although the rules are simple, there are 10^170 possibilities in Go. Therefore, it is very difficult to solve Go problems using enumeration. Previously, no computer program could beat a human player at Go.

The combination of deep learning and advanced search trees (reinforcement learning) adopted by AlphaGo creatively solves this problem in another way. It was also the unprecedented battle between AlphaGo and Lee Sedol that made DeepMind famous and made AI attract the attention of the world in an instant.

WaveNet

WaveNet is a text-to-speech model. Human speech is the result of multiple factors such as pitch, rhythm, and tone. Natural speech generation has always been difficult.

WaveNet is a generative model trained on speech samples. It creates waveforms of speech patterns by predicting which sounds are likely to follow each other. Each waveform is built one sample at a time, with up to 24,000 samples of sound per second. When WaveNet was released, its results were astonishing. Because this is a generative model released in 2016. This was a very cutting-edge approach at the time.

AlphaFold

AlphaFold is a deep learning model that can accurately predict three-dimensional models of protein structures and is accelerating research in nearly all areas of biology. Inside every cell in your body, billions of tiny molecular machines are hard at work. They enable your eyes to detect light, your neurons to fire, and the "instructions" in your DNA to be read, making you uniquely human.

These delicate, complex machines are proteins. They support not only the biological processes in your body, but every biological process in every living thing. They are the building blocks of life.

Currently, there are over 200 million known proteins, and more are discovered every year. Each one has a unique three-dimensional shape that determines how it works and what it does.

But figuring out a protein's exact structure remains an expensive, often time-consuming process and -- until now -- scientists have only been able to study the exact three-dimensional structure of a tiny fraction of the proteins known to science.

AlphaFold is a model DeepMind began working on in 2016. Its latest system can now predict the shape of proteins on a large scale in minutes, accurate to the atomic level. In 2021, Nature magazine detailed this research, which surprised everyone, and is open source on GitHub: https://github.com/deepmind/alphafold

AlphaTensor

DeepMind proposed AlphaTensor, the first artificial intelligence (AI) system for discovering new, efficient, and provably correct algorithms for basic tasks such as matrix multiplication. This sheds light on a 50-year-old open problem in mathematics: finding the fastest way to multiply two matrices.

Matrix multiplication is one of the basic operations of today's deep learning algorithms. Although it may have been taught in high school, it has a huge impact on today's digital world.

The operation of multiplying two matrices is used to process images on smartphones, recognize voice commands, generate graphics for computer games, run simulations to predict weather, compress data and video for sharing on the Internet, and much more. Companies around the world spend a lot of time and money developing computing hardware to perform matrix multiplication efficiently. Therefore, even small improvements in matrix multiplication efficiency can have widespread effects.

An AlphaTensor agent was trained to play the game using reinforcement learning, starting without any knowledge of existing matrix multiplication algorithms. Through learning, AlphaTensor gradually improved over time, rediscovering historically fast matrix multiplication algorithms such as Strassen's algorithm, eventually transcending the realm of human intuition and discovering algorithms that were faster than previously known.

Summary

In the early days, DeepMind had little profit pressure. It lost about US$680 million each year in 2018 and 2019. In 2019, Google also wrote off US$1.5 billion in debt. However, it may be due to Google's profit pressure or DeepMind losing too much, it began to try to make money with its own technology. The way it does this is by selling its technology to other companies and businesses owned by Google. In 2020, its revenue reached US$1.1 billion and it started to make money.

Compared with OpenAI, DeepMind's commercialization is still very restrained. Most of its results are still released as if it were a scientific research institution.