Opinion
Deep learning and neural networks have more to offer than their buzz suggests
Summary
- While Generative AI and its chatbots are powerful language-based technologies, it is still ‘good old AI’ that has been churning out astounding new products.
The computer that I am typing this article on is named Move37. My OnePlus Open phone is christened that too. Move37 is also the name of my fancy, app-controlled e-bike in Cambridge. They are all named after the magic 37th move made in a game of Go played between a human and an artificial intelligence (AI) engine in 2017, which opened my eyes to the power of AI. Lee Sedol, the undisputed world champion of Go—a complex, intuitive, ‘human’ game—was battling AlphaGo, an AI agent created by DeepMind (now Google DeepMind), perhaps the world’s premier Deep Learning company. Before the tournament, Sedol was confident to the point of dismissiveness, predicting that he will win the match by a landslide. In what is termed the ‘Sputnik Moment’ for AI, AlphaGo trounced him 4-1, eventually leading Sedol to retire from the game of Go in 2019. It was the 37th move with the 19th stone played by the AI in their second game that broke Sedol, a move described as “unique," and a move that “no human would ever make."
While it is ChatGPT and Generative AI that are making waves and dominating the discourse, I believe that the promised miracles of AI are happening elsewhere—using existing technologies like deep learning and neural networks. And, at the forefront of these life-changing discoveries is DeepMind, the company started by Demis Hassabis in London and now owned by Google. Hassabis demonstrated the power of this then-new technology through a demonstration of an AI agent acing the game Atari Breakout in 2013 after just 600 training sessions. Then came the epochal moment of the Go victory, which put deep learning in the spotlight. DeepMind took this further by creating AlphaZero, which could play any game at a beyond-human level. AlphaGo had been trained by analysing millions of Go moves, but the neural network powering AlphaZero was only given the rules of the game, and it learnt by playing against itself millions of times. It mastered chess in just nine hours, Shogi in 12 hours and then Go in 13 days.
Unlike AlphaGo, which learnt to play Go by analysing millions of moves from amateur games, AlphaZero’s neural network was only given the rules of each game. AlphaZero gave way to MuZero, which went one level further. It was not even told the rules of the game but became as good as AlphaZero in all games by learning a model of their environment, analysing the current position, the last action and deciding which action to take.
DeepMind and deep learning are just not fun and games, impressive as they are. Their next big breakthrough and the one that made me a convert was AlphaFold. This product cracked one of the hardest problems in medical science—predicting how a protein would fold. We, and every other carbon-based life form, are made of proteins, and it is how they fold that decides almost everything about our physiology and life. There are over 200 million known proteins today, and each of them folds in a unique three-dimensional shape. It was impossible for scientists to study each one of them, and that considerably hampered efforts to tackle disease. DeepMind started solving this problem right after the Sedol game, trained it on the sequences and structures of a hundred thousand proteins, and in four years brought it to a level where it can predict the folded shape of a protein, right down to the molecular level.
In 2020, this problem was declared solved. It was a breakthrough as important as mapping the human genome, or the discovery of antibiotics, something that can change medical science forever. As a scientist remarked, “What took us months and years to do, AlphaFold was able to do in a weekend."
DeepMind has kept its innovation factory going. Last December, it launched GNoMe, an AI tool that found more than 2 million new materials, of which nearly 400,000 could power future technologies around computer chips, solar panels and newer batteries. Google estimates that this is equivalent to 800 years’ worth of knowledge.
January 2024 saw the launch of AlphaGeometry, which could match and surpass the complex geometry problem solving skills of the planet’s brightest students. There are many more such wondrous innovations from DeepMind: SynthID for watermarking AI-generated images, Universal Speech Model which automatically recognizes all speech, AlphaCode, which writes advanced computer code, and Gemini, its Generative AI chatbot.
While GenAI and its chatbots are extremely powerful and promising language-based technologies, it is still ‘good old AI’ that has been churning out some astounding new products, led by DeepMind. Humans are still far ahead in general intelligence, but in these narrow niches, deep learning has achieved astonishing results.
In the Sedol-AlphaZero match, there was another vaunted move—No. 78 in the fourth game, this time by the human. Called the “God Move," it flummoxed and confused the AI, making it lose the game comprehensively to Sedol. However good AI becomes, we must always remember it is created by us.
