DeepMind

from Wikipedia, the free encyclopedia
DeepMind

logo
legal form
founding September 23, 2010
Seat London
management Demis Hassabis ( CEO )
Lila Ibrahim ( COO )
Number of employees 700
Branch Artificial intelligence
Website www.deepmind.com
As of December 1, 2017

DeepMind (formerly Google DeepMind ) is a company that specializes in programming artificial intelligence (AI). DeepMind was founded in September 2010 and acquired by Google Inc. in 2014 .

history

DeepMind Technologies was a British start-up founded in 2010 by Demis Hassabis , Shane Legg and Mustafa Suleyman. The first donors included venture capital firms Horizons Ventures and Founders Fund, and business angel Scott Banister. In 2014 the company received the “Company of the Year” award from the Cambridge Computer Laboratory.

On January 26, 2014, the US group Google announced the takeover of DeepMind Technologies and thus outperformed Facebook, which was also interested in DeepMind - it was the largest takeover by the Californians in Europe to date . The price remained a secret, industry experts assumed a purchase price of around $ 500 million (365 million euros). Several billionaires had already invested in DeepMind, including Elon Musk , CEO of the private space company SpaceX and electric car maker Tesla Motors , PayPal founder and original Facebook funder Peter Thiel , Skype co-founder Jaan Tallinn and Hong Kong magnate Li Ka-shing of Horizon Ventures , one of the most powerful men in Asia.

In 2014, DeepMind was named Company of the Year by the Cambridge Computer Laboratory .

After the takeover, the company was renamed Google DeepMind. Google also set up an ethics council to ensure that DeepMind's technology is not misused. The structure of the ethics council remained unclear.

In 2015, the company published the results of a research project in which Artificial Intelligence was supposed to learn old Atari games independently. According to Google DeepMind, the AI ​​succeeded in both learning the rules of the game and developing tactics for success on its own.

In December 2019, one of Suleyman's founders announced that he was leaving DeepMind so he could work at Google.

research

The official corporate goal of Google DeepMind is to understand intelligence ("Solve Intelligence"). In contrast to other Artificial Intelligences such as Deep Blue from IBM , Google DeepMind does not have a predefined target and is therefore more flexible in its application for various problems. Google DeepMind also differs in the basic structure of Artificial Intelligence. Instead of relying exclusively on a neural network , the AI ​​was expanded with a short-term memory in order to simulate the ability of an artificial memory. The developers of Google DeepMind therefore also refer to Artificial Intelligence as the “neural Turing machine ” and not as a neural network.

AlphaGo

AlphaGo was developed at DeepMind , a computer program that only plays the Go board game . In October 2015 it defeated the multiple European champion Fan Hui. It is the first program that was able to beat a professional Go player under tournament conditions. Between March 9th and 15th, 2016 AlphaGo competed against South Korean professional Lee Sedol , 9th Dan . The program won 4-1 after five rounds.

In 2017 an improved version called AlphaGo Zero was released, which beat AlphaGo 100 to 0, with AlphaGo Zero's strategies being created autonomously. The learning phase lasted only three days, compared to which AlphaGo took months to complete.

AlphaStar

In January 2019 AlphaStar was introduced, an AI program that plays the real-time strategy game StarCraft II . Like AlphaGo, it is an artificial neural network that first imitated human players and was then trained with reinforcement learning . AlphaStar won every game in two sets of five games each against the professional players Dario "TLO" Wünsch and Grzegorz "MaNa" Komincz. The number of actions per minute was limited to what is normal for humans. In contrast to human players, AlphaStar always has a complete overview of the visible parts of the map, but still only focuses on individual areas. MaNa won a game broadcast live at the presentation. AlphaStar's particular strengths were micromanagement and multitasking . In the last game she was thrown out of step by unusual actions by the human player.

WaveNet

WaveNet is a neural network that was originally developed from around 2014 for the purpose of improved, more natural-sounding text-to-speech synthesis (TTS), i.e. for more human-sounding reading systems, in which the program should analyze natural human voices, and so on to be able to generate more natural sounding language when reading texts aloud. Since March 2018, Google has been offering the WaveNet-based reading program Cloud Text-to-Speech as part of the Google Assistant .

In the meantime, DeepMind is also actively researching how to use WaveNet to reproduce existing individual human voices as precisely as possible, not just for reading texts aloud, in order to equip it with functions similar to those of the Adobe Voco introduced by Adobe Inc. in 2016 . A research report from June 2018 with the title Disentangled Sequential Autoencoder ("Unwoven sequential, automatic voice generator") found that WaveNet was successful in using WaveNet to replace the voice in an existing sound recording with any other real voice that contained the same text speaks ("content swapping"). To enable this ability, around 50 hours of recordings of the source and target voices are required, with which WaveNet must first learn the basic properties of the two voices (whereby it independently has its own model of the voice to be learned, abstracted from the voice recordings before it can do the voice conversion on any new sound recording of acceptable quality. The authors of the research report also emphasize that WaveNet is able to distinguish between static and dynamic properties (to: "disentangle"), ie the program automatically separates the properties to be retained during voice conversion (text content, modulation, speed, mood etc.) and the basic properties of the source and target voice to be converted.

In a follow-up report entitled Unsupervised speech representation learning using WaveNet autoencoders from January 2019, it was stated that DeepMind had further improved the distinction between static and dynamic voice characteristics in WaveNet. In the further follow-up report Sample Efficient Adaptive Text-to-Speech ("Sampling efficiency in adaptive text-to-speech synthesis") from September 2018 (last revision January 2019), DeepMind reports that the minimum number of voice recordings required for sampling a voice is meanwhile to have reduced a few minutes.

In 2014 already published report Towards end-to-end Speech Recognition with Recurrent Neural Networks ( "For the complete speech recognition means of recurrent neural networks ") Deep Mind pointed to successful attempts to use WaveNet automatic textualization of existing voice recordings.

The research report Large-Scale Visual Speech Recognition ("Generally applicable optical speech recognition") from July 2018 (last revision from October 2018) describes successful attempts to use WaveNet for lip reading with human lip movements also in completely silent video recordings, with the program being professional human Already clearly outperform lip readers in guessing the actual sounds. This feature seems to be developed as an independent WaveNet plug-in under the title LipNet .

Health data

In February 2016 the company announced that it was working with the UK's National Health Service to develop an iPhone app called Streams, but it was too early to say where AI was being used could. This should help to monitor patients who suffer from "kidney damage". April 2016, the specialist magazine " New Scientist " published details of the agreement, according to which Google will have access to the data of 1.6 million NHS insured persons who are annually in the three hospitals - Barnet, Chase Farm and the Royal Free - of the Royal Free NHS Trust treated. The far-reaching agreement allows data to be exchanged far beyond what has been announced. It contains data from HIV- positive patients, drug addicts or women who had abortions . There was no opt-out (possibility of refusal) for patients. In addition to the pathological and radiological results, the data of the intensive care medicine and the emergency departments are transmitted, as well as the complete daily activities of the clinics, the condition and accommodation of the patients and the hospital visits (who and when) are transmitted. In addition, having access to the central records of all NHS hospital treatments in the UK means that it can draw on historical data for the past five years - in addition to newly accruing data.

Critics fear for privacy and data protection . However, Google's staff should not be able to identify certain patients and the data should not be linked to Google accounts or products, said Mustafa Suleyman, head of the "Applied AI" division at DeepMind.

DeepMind Ethics and Society

In October 2017, DeepMind announced that it would found the DeepMind Ethics & Society research department, which will deal with the following topics: privacy, transparency and justice and the economic consequences thereof.

Web links

Individual evidence

  1. ↑ a b Google DeepMind company website .
  2. ^ "DeepMind Technologies Investors" (accessed March 12, 2015).
  3. ^ "Hall of Fame Awards: To celebrate the success of companies founded by Computer Laboratory graduates." from Cambridge University (accessed March 12, 2015).
  4. ↑ "DeepMind: Google buys start-up for artificial intelligence" , Spiegel.de, January 27, 2014
  5. ^ "Google Acquires Artificial Intelligence Startup DeepMind For More Than $ 500M," TechCrunch, accessed March 12, 2015.
  6. ^ "Google buys UK artificial intelligence startup Deepmind for £ 400m," The Guardian, accessed March 12, 2015.
  7. ↑ Google's super brain: DeepMind boss Demis Hassabis is building the ultimate thinking machine .
  8. ↑ January Samols: Hall of Fame Awards. January 23, 2018, accessed May 17, 2020 .
  9. ^ Google Beat Facebook for DeepMind, Creates Ethics Board .
  10. ^ "Google KI DeepMind learns old classic games independently" , Bluemind.tv, March 10, 2015.
  11. ↑ Subscribe to read | Financial Times. Retrieved May 17, 2020 .
  12. ↑ Playing Atari with Deep Reinforcement Learning , arxiv : 1312.5602v1
  13. ↑ "Google DeepMind: The Artificial Intelligence" from Bluemind.tv (accessed on March 12, 2015)
  14. ^ Neural Turing Machines , arxiv : 1410.5401
  15. ↑ a b Google software defeated Go-Genie in the last match. FAZ, March 15, 2016, accessed on September 13, 2017 .
  16. ^ Rory Cellan-Jones: Google DeepMind: AI becomes more alien . In: BBC News . October 18, 2017 ( bbc.com [accessed May 18, 2020]).
  17. ↑ James Vincent: Deep Mind's AI agents conquer human pros at Starcraft II. In: The Verge . January 24, 2019, accessed January 25, 2019 .
  18. ↑ Oord et al. (2016): WaveNet: A Generative Model for Raw Audio , Cornell University, September 19, 2016
  19. ↑ Novet, Jordan (2018): Google is finding ways to make money from Alphabet's DeepMind AI technology , March 31, 2018, CNBC
  20. ↑ Aharon, Dan (2018): Introducing Cloud Text-to-Speech powered by DeepMind WaveNet technology , Google Cloud Platform Blog, March 27, 2018
  21. ↑ Li & Mand (2018): Disentangled Sequential Autoencoder , June 12, 2018, Cornell University
  22. ↑ Chorowsky et al. (2019): Unsupervised speech representation learning using WaveNet autoencoders , January 25, 2019, Cornell University
  23. ↑ Chen et al. (2018): Sample Efficient Adaptive Text-to-Speech , September 27, 2018, Cornell University. Last revision of the same report from January 2019.
  24. ↑ Graves & Jaitly (2014): Towards End-to-End Speech Recognition with Recurrent Neural Networks , 2014, Cornell University
  25. ↑ Shillingford et al. (2018): Large-Scale Visual Speech Recognition , July 13, 2018, Cornell University
  26. ↑ Google's AI department goes to medical project - ZDNet.de . February 25, 2016.
  27. ↑ DeepMind is designed to help save lives .
  28. ↑ a b Hal Hodson: Revealed: Google AI has access to huge haul of NHS patient data .
  29. ↑ Laura Donnelly: Controversy as Google given access to NHS patient data . May 3, 2016.
  30. ↑ heise online: Great Britain: Google subsidiary DeepMind is allowed to evaluate millions of patient data .
  31. ↑ Warn data experts: Electronic health card completely unsuitable for the planned function - health . 4th May 2016.
  32. ↑ Why we launched DeepMind Ethics & Society. Retrieved May 18, 2020 .