As one of the largest artificial intelligence research facilities, DeepMind aims to advance the technologies of AI. DeepMind was acquired by large tech corporation Google for $600 million in 2014. Since the organization’s founding, it raked plenty of achievements shaping the world we live in today.
A Brief History of DeepMind
DeepMind is a British artificial intelligence research unit subsidiary of Alphabet Inc., founded in 2010. Founded by tech-experts Demis Hassabis, Mustafa Suleyman, and Shane Legg, they made it their mission to leverage AI in facilitating deep learning.
Following its success, Google acquired the organization for $600 million. Although Google became the parent company of DeepMind, Alphabet Inc. remained its subsidiary. There are multiple research centers all over the United States, Canada, and France under DeepMind Technologies.
DeepMind earned its reputation after introducing AlphaGo in 2016. This computer program utilizes AI to play the board game Go.
Subsequent versions of this program became so powerful that it beat a professional player and world champion in Go, Lee Sedol. The success of AlphaGo inspired the creation of another program called AlphaZero, which expanded on other board games, including chess, shogi, and Go Best.
After gaining a lot of traction, DeepMind received a lot of financial support from prominent venture capital firms such as Horizon Ventures and Founders Fund and influential entrepreneurs like Elon Musk and Scott Banister. After Google has taken over DeepMind, it often gets referred to as Google DeepMind.
Founding and early projects
AI researcher Hassabis started DeepMind with New Zealand-born ML researcher Shane Legg whom he had met at a computational neuroscience division of University College London (UCL). Completing the trio was Hassabis’s childhood friend Mustafa Suleyman.
The start-up was founded at a time when there was much less interest in AI than there is today. To accelerate the field, the team combined new ideas and advances from fields such as ML, neuroscience, mathematics, engineering, and computing and simulation infrastructure.
One of the earliest projects involved the teaching of AI tech to play 49 old and relatively primitive Atari games from the 70s and 80s such as Space Invaders. The AI was introduced to one game at a time and possessed no knowledge of the rules.
Over time, it learned how to play the games at an expert level just from seeing the pixels and score on the screen. In other words, the way a human player would. Ultimately, the project furthered DeepMind’s ambitions to create powerful, general-purpose AI that could be used in almost any situation.
Investment and Google acquisition
Early major investors included VC firms Horizons Ventures and Founders Fund in addition to entrepreneurs such as Elon Musk, Peter Thiel, Scott Banister, and Jaan Tallinn (who was also an adviser to the company).
On January 27, 2014, DeepMind was acquired by Google for £400 million – the company’s largest European purchase to date. Around nine months later, DeepMind won the Company of the Year award from the University of Cambridge.
Partnership with NHS
DeepMind entered into a collaboration with the Royal Free Hospital in 2015 – a major teaching hospital in London. The end result was a patient safety app called Streams that could identify sickness warning signs from test results and alert hospital staff if an assessment was required.
While the partnership saved nurses and clinicians around two hours per day, it was ruled in 2017 that Royal Free Hospital had breached the Data Protection Act when it shared the personal data of 1.6 million patients with DeepMind.
In 2016, the DeepMind AlphaGo program beat world Go champion Lee Sedol four times out of five. While machines had been able to beat humans at chess and even in Jeopardy contests, the Chinese board game Go was a different story.
Compared to 20 possible moves in an average position in chess, there are 200 possible moves in Go which yield more permutations than there are atoms in the universe. AlphaGo learned the game by analyzing the historical data of hundreds of previous games and the moves Go champions made to win. Based on this analysis, the AI then devised its own strategies.
Various improvements were made to the AI over the following years. Astoundingly, an updated version called AlphaGo Zero was able to defeat the old version AlphaGo 100 games to 0.
DeepMind’s Exploration to Machine Learning
The primary goal of DeepMind is to leverage the advanced disciplines of machine learning and neuroscience to build general-purpose learning algorithms. The organization has a belief that these algorithms are not only the key to improve AI but to understand the human mind better. Since the organization started, it has been releasing publications surrounding the topic of artificial intelligence. A year after Google’s acquisition of DeepMind, it released an open-source testing tool called GridWorld. This program evaluates the behavior of specific algorithms under particular circumstances. This testbed has a kill switch when undesired behavior gets detected from the algorithms to promote AI safety while exploring the discipline’s capacity.
In 2018, the company started developing systems that can compete in a variety of games. The system began playing the 1999 multiplayer-focused first-person game called Quake III Arena. It allows them to test the capabilities of the systems as several games modify their behaviors. Leveraging machine learning in strategic games, including chess, hones the critical thinking capabilities of computer systems. They can retain what they have learned upon being immersed in complex games. Simply put, gamification’s objective is to let the system learn to acquire human-like intelligence and behavior.
Transitioning Towards Reinforcement Learning
As the research advances towards machine learning, DeepMind began exploring more of its aspects. Taking deep learning to a whole different level, researchers studied deep reinforcement learning. Unlike other artificial intelligence systems that most of us are familiar with, it makes decisions and executes actions independently. In other words, reinforcement learning systems are not pre-programmed. The system solely relies on experience to learn and develop, just like human beings.
DeepMind tests the capability of systems by letting them play the game without programming them with its instructions. They need to work their way towards the top after going through multiple matches against a human opponent. After a couple of attempts, systems often become more knowledgeable of the gameplay and become much better than their opponents. The goal of leveraging reinforcement learning is to eliminate human error from affecting the gameplay’s efficiency level. Since its introduction, many researchers have tested its potential for programs other than games.
DeepMind’s Contribution to the Healthcare Industry
One of the most notable contributions of the organization is its collaboration with WaveNet. DeepMind established WaveNet to advance healthcare services and promote patient care. The software generates speech that emphasizes human-like features. WaveNet offers text-to-speech systems that are less robotic and natural-sounding to assist people who have a speech impairment.
The organization coordinated with individuals who have suffered from the same condition as the program’s target market. They worked with former NFL player Tim Shaw, who has Amyotrophic Lateral Sclerosis (ALS). DeepMind created a system that produces the natural voice as if it was them uttering the words. Other programs may have required the patients to spend a long time recording audios and reading scripts. Unfortunately, speech-impaired patients do not have the luxury to record audios to have the system recreate their voice. WaveNet only needs a few audio recordings, and the algorithm will take care of the rest. It took them six months to perfectly recreate the voice of Tim Shaw. The text-to-speech program presents Shaw’s voice before his disease.
Developments of DeepMind Under Google
Following the acquisition of Google to the notable AI research organization, it brought several changes to the platform’s operations. DeepMind influenced the AI department of the popular search engine website in a few ways:
- One of the most current use-cases of DeepMind is the advancement of application recommendations within the Google platforms. Algorithms get leveraged to make the feed more personalized for the user. All they had to do is to gather data based on the user preferences, app behavior, and actions. This use allows the users to quickly find the applications that they will most likely find useful.
- Secondly, another major project involving DeepMind is algorithms’ production to cool down the servers across Google platforms. These features allow better usage and more efficient navigation around the application. Offering adaptive brightness and optimized battery, browsing through Google platforms can be a breeze. You can leverage machine learning to conserve the energy of mobile devices, laptops, and computers. From the app itself, users can modify lighting under certain conditions.
Overall, continuing the evolution of DeepMind allows Google to improve the user experience. When they create a platform that everyone finds easy to use, more people can trust and rely on it. With the advancements of DeepMind, it seems Google has the opportunity to make nearly anything user-friendly.
- DeepMind is a British AI company and research laboratory that is now part of Alphabet, Inc. It was founded in 2010 by Demis Hassabis, Shane Legg, and Mustafa Suleyman, and is now headquartered in London.
- DeepMind was started at a time when there was much less interest in AI than there is today. One of the earliest projects involved the teaching of AI tech to play 49 old and relatively primitive Atari games from the 70s and 80s such as Space Invaders.
- DeepMind entered into a collaboration with the Royal Free Hospital in 2015 to build AI that could scan patient records and alert staff to sickness. However, it was shut down because of privacy concerns. DeepMind’s AlphaGo program was also developed around this time and managed to beat a human player (and then itself) at the Chinese board game known as Go.
Connected AI Concepts
Deep Learning vs. Machine Learning
OpenAI Organizational Structure
Stability AI Ecosystem
Main Free Guides: