Google DeepMind

DeepMind is creating the Streams application, with which it will be possible to identify patients who may develop acute kidney failure, and this company needed information about real patients. The organization received information about more than one and a half million patients, including information such as experience with drugs, abortion, HIV status.

Performance indicators

2019: $1.5 billion in losses and debt relief

In 2019, DeepMind's net losses reached £477 million, 1.5% higher than the cash losses of a year ago. This is stated in the statements that Google disclosed in the register of legal entities in the UK (Companies House).

DeepMind's revenue at the end of 2019 turned out to be £266 million, significantly increasing compared to a year ago at £103 million. However, major revenues come from Google and other Alphabet holding companies that pay DeepMind for research and development.

Google wrote off $1.5 billion in debt to its unprofitable AI division

The vast majority of DeepMind's costs in 2019 came from "staff and other related costs." The annual report shows around £468m has been spent on these targets, up from £398m in 2018. Administrative costs, including staff and infrastructure costs, rose 26% to £717m.

According to reports, Google wrote off DeepMind debts in the amount of £1.1 billion (about $1.5 billion), and also intends to continue to provide financial support to its AI division, despite the growing losses in it.

I am delighted with the pace with which our engineering and research teams work at both Google and DeepMind, "said Alphabet CEO Sundar Pichai.

Google's desire to help DeepMind is largely due to the success of this division. So, in the fall of 2020, it achieved great achievements in protein folding. The latest version of the AlphaFold AI system is capable of providing reliable predictions of the basic physical structure of the protein in a matter of days. The breakthrough was recognized by the organizers of the critical assessment of protein structure prediction (CASP). By December 2020, DeepMind employs about 1000 people, including leading research scientists in the field of artificial intelligence with annual salaries of $1 million.^[1]

History

2024: Developing a robot for playing table tennis

In early August 2024, Google DeepMind introduced a robot that plays table tennis better than humans at the amateur level. The developers claim that this is the first time that the robot has been taught to play with people in active sports.

According to the test results, a robotic arm with a 3D printed racket was able to win 13 out of 29 rounds of table tennis against human rivals of various levels. However, the system is still far from perfect, as it has never been able to defeat an advanced player.

GOOGLE DEEPMIND

According to the developers, this system is a new step towards creating robots that can perform various useful tasks in real conditions, for example, at home and in warehouses. Google DeepMind's approach to machine training is applicable in many other areas, said computer systems researcher Lerrel Pinto. The Google DeepMind researchers used a two-step approach to train the simulation system: They used computer simulations to teach the system strike skills and then tuned it using real-world data, allowing the system to improve over time.

During matches with humans, the robot collects data on its performance to hone skills. It tracks the position of the ball using data from a pair of cameras and monitors its human opponent's playing style with a motion capture system using LEDs on the opponent's racket. Ball data is returned to the simulation model for further training, creating a continuous feedback loop.

This feedback allows the robot to test new skills, meaning it can adjust its tactics and behavior in the same way a person does. This means that he will play better and better both throughout a particular match and in general. It is assumed that in the future such reinforcement training could help robots cope with new tasks that they have not tried before.^[2]

2019: DeepMind helped Google boost performance of its wind farms by 20%

On February 27, 2019, it became known that researchers from the DeepMind laboratory taught the neural network to predict the performance of wind turbines in 36 hours based on retrospective data on their performance and weather in a certain area. According to the researchers, the use of such an algorithm has improved the performance of Google's wind farms by 20 percent. Read more here.

2018

Google joins DeepMind medical project to create AI assistant for carers and doctors

On November 13, 2018, as part of an ongoing reorganization of efforts in the field health care Google , it announced the joining of a medical project, DeepMind Health which was previously part of the research center of AI the London-based company DeepMind, which is part of the holding along with Google. Alphabet More. here

DeepMind Fully Automates Google Data Center Cooling Process

In August 2018, it became known about the complete automation of the cooling process. data centers Google Thanks to this, the company was able to reduce energy consumption by 30%, according to Google startup DeepMind, which has developed a new control system. More. here

2016

Google has opened free access to its artificial intelligence

Google DeepMind - Alphabet Inc.'s artificial intelligence development (AI) arm - has opened its platform. The source code of the DeepMind Lab virtual environment (formerly Labyrinth) for machine learning will be completely and free for users on the basis of the largest web service for hosting IT projects and their joint development GitHub^[3]

According to Bloomberg, now anyone can download the code and configure it in order to train their own artificial intelligence systems. Developers will also be able to create new game levels for DeepMind Lab and upload them to GitHub.

Journalists note that the disclosure of the source code continues Alphabet's policy aimed at maximizing the openness of its scientific research. Last month, the company signed an agreement with developers from Activision Blizzard Inc. to use the popular Starcraft II game to train artificial intelligence.

The task of forming radiation maps in radiotherapy of cancer tumors

In September 2016, it became known that Google DeepMind received two medical tasks for diagnosis: to determine renal failure and senile retinal macular degeneration. Now he begins to help doctors and in treatment. More precisely, in the formation of a radiation map during radiotherapy of cancer tumors in the head and neck, which is created according to the results of tomography. The fact is that the radiation map in these areas should be super accurate and an experienced specialist needs at least 4 hours of routine work to draw it up. So far, artificial intelligence will not be fully trusted in such an important matter, but the idea is that after his work, the doctor will already need only 1 hour to draw up a map. And this should reduce the cost and speed up the treatment of many people^[4].

Creation of DeepMind Health Division

On February 24, 2016, the creation of the DeepMind Health division was announced, which was headed by Mustafa Suleiman, one of the founders of the company.

It was reported that in its work, DeepMind Health will cooperate with the National Service, health care Great Britain provide doctors with technical expertise to develop and improve patient care technologies. Special attention is planned to be paid to the security of patient data and confidentiality. The unit's activities should be reviewed by a board of independent experts, including Richard Horton, editor of the authoritative magazine on to medicine The Lancet.

At this time, DeepMind Health is working to create electronic tools that make it easier for a doctor. A smartphone app has been introduced to more accurately diagnose acute kidney failure. Also, DeepMind bought a medical application, a taxi manager for doctors. The team of doctors from Imperial College London that created it joins DeepMind Health. The amount of the transaction was not disclosed.

2015: AlphaGo System Success

The company gained fame as a result of the development of the AlphaGo system - it defeated a professional go player. DeepMind has created a neural network capable of learning video games and playing at the level of human capabilities.

2014: Google buys the company. Rename to Google DeepMind

On January 26, 2014, Google announced the purchase of DeepMind Technologies. Various sources reported: the transaction amount ranged from $400 to $650 million.

2010: Founding of DeepMind Technologies

The company was founded in 2010 in London under the name DeepMind Technologies. The company was created by Demis Hassabis, Shane Legg and Mustafa Suleyman. Before that, Hassabis and Legg were already familiar from University College London, where they worked in the Gatsby Computational Neuroscience Unit.

The company has invested large venture capital funds Horizons Ventures, Founders Fund, as well as entrepreneurs Scott Banister and Elon Musk. Jan Tallinn was one of the company's early investors and advisers.

Notes