Top DeepMind AI Products Revolutionizing The World

When DeepMind launched in 2010, there was little interest in the field of artificial intelligence (AI) compared to the levels of interest that exist today. To accelerate the nascent technology field, the team adopted an interdisciplinary approach.

They integrated new ideas with advances in engineering, machine learning, simulation and computing infrastructure, neuroscience, mathematics, and new methods of organizing scientific endeavors.

DeepMind Technologies is a British artificial intelligence subsidiary of Alphabet Inc. The London-based research laboratory was acquired by Google in 2014. This firm has research centers in France, Canada, and the United States. In the next year, it became entirely owned by Alphabet.

The firm joined forces with Google to accelerate its work and continued to set its research agenda. Several of the DeepMind programs have learned to diagnose eye diseases as efficiently as the world’s top doctors and to save 30% of the energy that is used to ensure data centers remain cool. The programs predict the complex 3D shapes of proteins that could transform how drugs are invented in the future.

The company achieved early success in computer games with researchers normally using it to test AI. One of the programs learned to play 49 different Atari games from scratch, just from seeing pixels and scores on the screen. The AlphaGo program was also the first one to beat a professional Go player, a feat that is described as a decade ahead of its time.

Over the years, DeepMind created a neural network that learns how to play video games like humans, and a Neural Turing machine, or a neural network that can access an external memory just like the conventional Turing machine. The development resulted in a computer that mimics the near-term memory of the human brain.

In 2016, DeepMind made headlines after its AlphaGo program managed to beat a human professional Go player Lee Sedol, the world champion, in a 5-game match, that became the subject of a documentary film.

Another general program, AlphaZero, beat the most powerful programs playing chess, Go, and Shogi (Japanese Chess) after several days of playing against itself using some reinforcement learning. In 2020, DeepMind made considerable advances in the protein folding problem.

DeepMind Overview

Demis Hassabis, Shane Legg, and Mustafa Suleyman are the founders of this thriving company. Legg and Hassabis first met at University College London’s Gatsby Computational Neuroscience Unit.

Initially, the company started working on artificial intelligence technology teaching it to play some old games from decades earlier.

Some of the games included Space Invaders, Pong, and Breakout. The developers introduced artificial intelligence to one game at a time without having any previous knowledge of its rules. After the technology spent some time learning how the game operates, AI would  then go on to become an expert in it:

“The cognitive processes which the AI goes through are said to be very like those a human who had never seen the game would use to understand and attempt to master it.”

The founders aimed to create a general-purpose artificial intelligence that can be used effectively and efficiently for almost anything. Horizons Ventures and Founders Fund are some of the main ventures that invested in the company. Also, notable entrepreneurs like Peter Thiel, Scott Banister, and Elon Musk invested in the company during its early days.

On January 26, 2014, Google acquired DeepMind for $500 million in the same year when it received the Cambridge Computer Laboratory “Company of the Year” award. The sale to Google came after Facebook ended its negotiations with the company in 2013. Afterward, the company was rebranded as Google DeepMind and maintained the name for two years.

The Royal Free NHS Trust and DeepMind signed their first Information Sharing Agreement (ISA) in September 2015 to create Streams, a clinical task management app. After acquisition by Google, the firm established an AI ethics board for research but it remains a mystery with both companies declining to say who sits on the board.

The company joined Facebook, Amazon, Microsoft, Google, and IBM to launch ‘Partnership on AI’ devoted to the society-AI interface. DeepMind has opened a new unit known as DeepMind Ethics and Society focusing mainly on the ethical and societal questions that are raised by AI technology. The prominent philosopher, Nick Bostrom, is an advisor of the ‘Society.’

DeepMind Products And Technologies

The company strives to integrate the best techniques from systems neuroscience and machine learning to create a powerful general-purpose learning algorithm. In 2016, Google Research published a paper about AI Safety and how to avoid undesirable behavior during the artificial intelligence process.

In 2017, DeepMind released GridWorld, which is an open-source testbed for evaluating whether an algorithm learns to disable the kill switch or exhibits some undesirable behaviors. Sometime in July 2018, the researchers at the company trained one of its systems to play the Quake III Arena computer game.

As of last year, the firm had published more than a thousand papers, with 13 of these papers being accepted by Science or Nature. Here are some of the top DeepMind products.

Deep Reinforcement Learning

As opposed to the other AIs that were developed for pre-defined purposes and function within a limited space, DeepMind says that its system is not pre-programmed. The technology learns from experience by utilizing just raw pixels as data input.

It mostly uses deep learning running on a convolutional neural network using a novel type of Q-learning. Q-learning is a type of model-free reinforcement learning. The technology tests the system on video games, including early arcade games like Breakout and Space Invaders.

Then, without changing the code, the AI system starts to understand how to play the game and after playing a few sessions, it plays more efficiently than any human. Back in 2013, DeepMind posted in-depth research on an AI system that can surpass human abilities in various games, leading to its acquisition by Google.

Last year, the company unleashed Agent57 and artificial intelligence Agent that exceeds human-level performance on all the 57 games of the Atari2600 suite.

AlphaGo And Successors

In 2014, the firm published research on computer systems with the ability to play the Go game. Later in October 2015, AlphaGo, a computer Go program, developed by the company beat the European Go champion Fan Hui, five to zero. That was the first time when an AI program defeated a professional Go player.

In March 2016, the AlphaGo beat Lee Sedol, one of the highest-ranked players worldwide, with a score of 4-1. During the 2017 Future of Go Summit, the AI won a 3-game match with the world number 1 at the time, Ke Jie. The system used a supervised learning protocol, studying many games played by humans against one another.

The improved AlphaGo Zero version defeated the previous AlphaGo system 100 games to 0 in 2017. The newer version’s strategies were self-taught and it beat its predecessor within three days with less processing power than AlphaGo. Later in the year, a modified version of AlphaGo Zero, AlphaZero gained superhuman abilities at shogi and chess.

All these versions of DeepMind’s artificial intelligence systems learned playing only through self-play. AlphaGo technology was designed to use the deep reinforcement learning approach enabling it to improve over time through self-learning.

The system used two deep neural networks enabling it to evaluate move probabilities and a value network to assess positions. This policy network was trained through supervised learning and was then refined by policy-gradient reinforcement learning. In that context, the value network learned to determine winners of the games played by the policy network against itself.

Later, the network used a lookahead Monte Carlo tree search (MCTS) that used a policy network to determine candidate high-probability moves as the value network simultaneously evaluated tree positions. The system was using reinforcement learning where the system played millions of these games against itself aiming to increase its win rate.

Notably, its simplified tree search relies mainly on its neural network to evaluate the positions and sample moves without using the Monte Carlo rollouts. With these enhancements, the AlphaZero system needed less computing power than AlphaGo, operating on four specialized AI processors known as Google TPUs instead of the 48 used by AlphaGo.

AlphaFold

Sometime in 2016, DeepMind turned its artificial intelligence research and development to one of the toughest challenges existing in science, protein folding. Barely two years later, DeepMind’s AlphaFold was awarded the 13th Critical Assessment of Techniques for Protein Structure Prediction (CASP) trophy after it successfully determined the most accurate structure for 25 out of 43 proteins.

Hassabis commented in an interview with The Guardian:

“This is a lighthouse project, our first major investment in terms of people and resources into a fundamental, very important, real-world scientific problem.”

Last year, during the 14th CASP, AlphaFold’s projections got an accuracy score comparable with lab techniques. One member of the panel of scientific adjudicators, Dr. Andriy Kryshtafovych, said that the achievement was ‘truly remarkable, and added that the problem of predicting how the proteins fold had been extensively solved.

Other Notable DeepMind Products

The company introduced a text-to-speech system, WaveNet, in 2016. At first, it was too computationally intensive for use in consumer products but it became ready for use on applications like Google Assistant in late 2017. In the following year, Google unveiled Cloud Text-to-Speech, a commercial text-to-speech product, based on WaveNet.

Later in 2018, DeepMind developed a highly efficient model known as WaveRNN co-developed using Google AI that was rolled out to Google Duo users in 2019.

Google says that the DeepMind algorithms have majorly increased the efficiency of cooling most of its data centers. Also, the technology assists Google Play’s personalized app recommendations and collaborated with the Android team to create a pair of features made available to the Android Pie devices.

The new features include Adaptive Brightness and Adaptive Battery which use machine learning to save energy and make devices running the operating system more user-friendly. That was the first time that DeepMind integrated these techniques in a small scale magnitude with the normal machine learning applications needing a lot of computing power.

The company’s Hubble telescope enabled people to look deeper into space, with the tools available already expanding human knowledge and, in turn, making a positive global impact. DeepMind’s long-term mission is to solve intelligence, creating generalized and effective problem-solving systems, dubbed artificial general intelligence (AGI).

Entirely guided by ethics and safety, the invention may be held the society to get viable solutions to some of the most challenging and fundamental scientific issues in the world.

For now, the company keeps developing its technology and it is aiming to expand its usability in almost all critical facets of humanity including health, gaming, and environmental preservation.

Source: https://e-cryptonews.com/deepmind-ai-products/