AI wins a contest --8/7/24

07 Aug 2024

AI wins a contest --8/7/24

0 COMMENTS

Today's selection -- from Genius Makers by Cade Metz. In 2012, a student named George Dahl used neural networks, a type of artificial intelligence, to pioneer a new approach to discovering new drugs:

“Three years earlier, in the summer of 2012, Merck &Co., one of the world's largest drug companies, launched a contest on a website called Kaggle. Kaggle was a place where any company could set up a contest for computer scientists, offering prize money to anyone who could solve a problem it needed solved. Offering a $40,000 prize, Merck provided a sprawling collection of data describing the behavior of a particular set of molecules and asked contestants to predict how they would interact with other molecules in the human body. The aim was to find ways of accelerating the development of new medicines. Two hundred thirty-six teams entered the contest, which was scheduled to run for two months. When George Dahl, Geoff Hinton's student, discovered the contest while riding on a train from Seattle to Portland, he decided to enter. He had no experience with drug discovery, just as he had no experience with speech recognition before building a system that shifted the future of the entire field. He also suspected that Hinton wouldn't approve of him entering the contest. But then Hinton liked to say he wanted his students working on something he wouldn't approve of ‘It is sort of like the Godel completeness result. What if he approves of you doing things he doesn't approve of? Is that really disapproval?’ Dahl says. ‘Geoff understands the limits of his own abilities. He has some intellectual humility. He is open to surprises, to possibilities.’

“When Dahl returned to Toronto, he met with Hinton, and when Hinton asked, ‘What are you working on?,’ Dahl told him about Merck.

“‘I was on the train going to Portland and I just trained a really dumb neural net on this Merck data and I hardly did anything at all and it's already in seventh place,’ Dahl said.

“‘How much longer is the contest?’ Hinton asked.

“‘Two weeks,’ Dahl said.

“‘Well,’ Hinton replied, ‘you have to win it.’

“Dahl wasn't really sure he could win it. He hadn't given the project all that much thought. But Hinton was insistent. This was the heady time between the success of deep learning with speech technology and the triumph with Go, and Hinton was keen to show just how adaptable neural networks could be. He now called them dreadnets (a play on the early-twentieth-century battleships called dreadnoughts), convinced they would sweep everything before them. Dahl was reminded of an old Russian joke Ilya Sutskever liked to tell in which the Soviet army runs out of shells while firing on its capitalist enemy. ‘What do you mean you're out of shells?’ the Soviet general says to the sergeant who alerts him to the problem. ‘You're a communist!’ So the army keeps on firing. Hinton said they had to win the contest, so Dahl enlisted the help of Navdeep Jaitly and a few other deep learning researchers from their lab in Toronto—and they won.

“The contest explored a drug discovery technique called the quantitative structure-activity relationship, or QSAR, which Dahl had never heard of when he went to work on the Merck data. As Hinton put it: ‘George wiped out the whole field without even knowing its name.’ Soon, Merck added this method to the long and winding processes necessary to discover new medicines. ‘You can think of AI as a large math problem where it sees patterns that humans can't see,’ says Eric Schmidt, the former Google chief executive and chairman. ‘With a lot of science and biology, there are patterns that exist that humans can't see, and when pointed out, they will allow us to develop better drugs, better solutions.’

“In the wake of Dahl's success, countless companies took aim at the larger field of drug discovery. Many were start-ups, including a San Francisco company founded by one of George Dahl's labmates at the University of Toronto. Others were pharmaceutical giants like Merck that at least made loud noises about how this work would fundamentally change their business. All were years away from overhauling the field completely, if only because the task of drug discovery is prodigiously difficult and time-consuming. Dahl's discovery amounted to a tweak rather than a transforming breakthrough. But the potential of neural networks quickly galvanized research across the medical field.”