More than 13,000 artificial intelligence maves flocked to Vancouver this week for the world’s leading academic AI conference, NeurIPS. The location included a maze of colorful company cabins to attract recruits for projects such as doctor’s software. Google handed out free baggage trays and socks with the colorful bicycles that employees ride on campus, while IBM offered hats with “I ❤️A????”.
On Tuesday evening, Google and Uber organized well-run parties with excess parties. The next morning at 8.30 am, one of the best Google researchers gave a keynote with a sobering message about the future of AI.
Blaise Aguera y Arcas praised the revolutionary technique known as deep learning, allowing teams like his phones to recognize phones and voices. He also regretted the limitations of that technology, designing software called artificial neural networks that could become better in a specific task through experience or seeing labeled examples of correct answers.
“We are a bit like the dog that took the car,” said Aguera y Arcas. Deep learning has quickly brought down a number of long-term challenges in AI – but does not immediately seem suitable for many who remain. Problems that involve reasoning or social intelligence, such as weighing a potential appointment in the way a person would act, are still out of reach, he said. “All the models we have learned to train are about passing a test or winning a game with a score [but so many things] that intelligences do not fall under that heading,” he said.
Hours later, one of the three researchers who were the godfathers of deep learning also pointed to the limitations of the technology he had helped bring into the world. Yoshua Bengio, director of Mila, an AI institute in Montreal, recently shared the highest prize in computer science with two other researchers for starting the deep learning revolution. But he noted that the technique produces highly specialized results; a system that is trained to show superhuman performance in one video game is unable to play the other. “We have machines that learn in a very scary way,” said Bengio. “They need much more data to learn a task than human examples of intelligence, and they still make stupid mistakes.”
Bengio and Aguera y Arcas both urged NeurIPS participants to think more about the biological roots of natural intelligence. Aguera y Arcas showed the results of experiments in which simulated bacteria were adapted to search for food and to communicate via a form of artificial evolution. Bengio discussed early work to make profound learning systems flexible enough to handle situations that are very different from those they were trained in, and made an analogy with how people can handle new scenarios, such as driving in another city or country.