DeepMind

Published by Daniel at February 1, 2022

DeepMind Tech has revolutionized deep learning by infusing "creativity" to master ancient games and forming partnerships with Google and healthcare...

DeepMind Technologies is a United Kingdom-based company devoted to AI development and one of the founding members of the Partnership on AI to Benefit People and Society. The origins of the company date back to 2010 when it was founded by Shane Legg, Demis Hassabis, and Mustafa Suleyman to accelerate the Artificial Intelligence field when it was unknown, feared, and unexplored. This company’s approach was that the best way to explore the unknown was using a multidisciplinary team that included engineering, mathematics, neuroscience, and computing infrastructure. While developing machine learning models, they tested their results using computer games (a common way to test AI performance), obtaining neural networks that learned how to perform remarkably. Results were so promising that there were extensive negotiation attempts between Google and Facebook, but Google was later restructured as Alphabet Inc and acquired the company in January 2014 .(1,2,3)

Mastering the game

The product of this alliance boosted DeepMind’s project development. It expanded the company, resulting in the acquisition of a much more challenging task, mastering the ancient Chinese game of Go, considered the most challenging game for AI development. While machines defeating world champions of chess and backgammon were accomplishments in the 1990s, the Go game was, by far, more complex with ten times more probabilities per move than chess and around 10 to the power of 170 possible board configurations. DeepMind developed the program AlphaGo and trained its deep neural network using matches against amateur players first and matches against different versions of itself for thousands of cycles.(4)

DeepMind developed the program AlphaGo and trained its deep neural network using matches against amateur players first and matches against different versions of itself for thousands of cycles. (4)

AlphaGO

After its training, AlphaGo subsequently defeated the European Go Champion Mr. Fan Hui (2015), the legendary owner of 18 world Go titles Mr. Lee Sedol (2016), and later the World’s Champion of Go Mr. Ke Jie (2017). AlphaGo Zero was developed by training a new and “naive” AlphaGo network by playing against itself, starting with a random play, and then against the fully developed AlphaGo. The results surpassed the game’s mastery in previous versions by enhancing the program’s creativity with initial random playing and subsequently polishing its skills with human-known techniques. (4)

The results surpassed the game’s mastery in previous versions by enhancing the program’s creativity with initial random playing and subsequently polishing its skills with human-known techniques. (4)

One for all, all for one

An important and contrasting feature of DeepMind neural networks is that the company’s perspective on AI is that Deep Neural Networks should not be pre-programmed with a specific function. This approach makes the core or backbone of these networks applicable to several purposes and the uses of AI, endless. They keep developing projects based on the premise that “Artificial intelligence should be general.”(5)

Foreseeing the future, one vision at a time

Notable projects in the health area include a partnership with Moorfields eye hospital for the implementation of Artificial Intelligence in analyzing complex images used in ophthalmology (e.g., optical coherence tomography) for the detection of diabetic retinopathy or age-related macular degeneration at such a degree; it can predict worrisome progressions of age-related macular degeneration even months before it happens.(6,7)

Unraveling the Levinthal’s Paradox

Another promising field consists of AI applications to predict tridimensional structures of proteins based on their amino acid sequence. For several decades scientists have struggled to determine and predict both the structure and physical properties of proteins based on their sequence. Whereas sometimes possible to determine the structure of folded proteins using costly methods such as cryo-electron microscopy, nuclear magnetic resonance, or X-Ray crystallography, these methods involve trial and error and cannot predict structures of unfolded amino acid sequences. Predicting protein structure may enhance drug development, genetic therapies, and protein/enzyme engineering in unprecedented ways. In 2019 the company discovered a new area in deep learning called distributional reinforcement learning. Scientists believe this pathway has important implications for neuroscience and could enlighten researchers on how the dopaminergic reward system works and understanding motivated goal-directed behavior.(8,9)

"Googling" Deep into the Mind"

By sharing Alphabet, Inc as their parent company, DeepMind and Google work closely. Part of the AI projects developed by DeepMind has improved Google products and infrastructure. A daily use example includes the implementation of WaveNet in the Google Assistant, a Deep Neural Network capable of processing and producing a more human-sounding speech for Google products and enhancing their speech-processing speed. This project gained importance for many individuals presenting with speech impairment, including patients with Amyotrophic Lateral Sclerosis.

The algorithm developed by DeepMind requires a few audio recordings to recreate voice, compared with previous systems. Another exceptional advent is the AI trained optimization of Google servers cooling systems that reduced by 30% the energy associated with its maintenance and its corresponding carbon emissions and this collaboration has also been prolific to Google Maps, where AI has produced a significant reduction in the percentage of inaccuracies in the time of arrival estimates by 51%.(10,11,12)

This project gained importance for many individuals presenting with speech impairment, including patients with Amyotrophic Lateral Sclerosis.The algorithm developed by DeepMind requires a few audio recordings to recreate voice, compared with previous systems. Another exceptional advent is the AI trained optimization of Google servers cooling systems that reduced by 30% the energy associated with its maintenance and its corresponding carbon emissions and this collaboration has also been prolific to Google Maps, where AI has produced a significant reduction in the percentage of inaccuracies in the time of arrival estimates by 51%.(10,11,12)

Multidisciplinary, multicultural, and multipotential

DeepMind is also committed to potentiating education and promoting diversity by providing scholarships for M.Sc. and Ph.D. in computer science and data analysis and mentoring students from underrepresented populations and diverse cultural backgrounds to incur in the AI field. Therefore, nourishing future assets of AI and the company.(13)

Although its headquarters is located in London, DeepMind also counts with global laboratories in Canada (Alberta and Montreal), the United States of America (Mountain View), and France (Paris) with a varied and inclusive community and, most importantly, different perspectives to tackle the most challenging projects.(14)