The enormous potential of artificial intelligence has electrified industries and scientists eager to use these new technologies for commercial applications and discovery. That enthusiasm brought representatives from the private sector, government, and academia to Chicago earlier this month for the U.S. Department of Energy (DOE) Innovation XLab Artificial Intelligence Summit, a two-day event featuring speakers, panels, and demonstrations about using AI in healthcare, national security, transportation, and many other areas.
Rebecca Willett, Professor of Statistics and Computer Science at the University of Chicago, shares that excitement about AI and machine learning. But in her keynote address on the second day of the summit, she also detailed the many challenges that remain before AI can be reliably deployed for these applications — and the critical role for partnerships between industry, government, and academia in overcoming those hurdles.
In her talk, Willett first covered the many different applications that could soon benefit from the integration of AI tools, including better forecasts for weather and agriculture, optimization of energy usage and transportation in cities, and improved diagnosis and treatment of disease.
“It’s hard to imagine any aspect of our lives that’s going to remain untouched by artificial intelligence,” Willett said in her keynote, which is partially available here (skip to 1:02:00).
But Willett also stressed that we are still in the earliest stages of AI applications, and that there is much more work to be done on the foundations of these computational tools. Technologies, such as computer vision, that have achieved major advances in the past decade remain vulnerable to small perturbations — changing a few pixels can trick a system into mislabeling a panda as a gibbon, while small natural motions of a polar bear can dramatically change how well it can be identified.
The issue, Willett said, is that many of today’s AI systems “learn from examples” — the data that is fed into them — and too few examples, or examples that are insufficiently diverse, can have unpredictable impacts on how AI systems perform The experimentalist approach of trying different ideas out to see what works is excellent in low-stakes settings such as identifying animals in images on the internet, but in many emerging application domains, mistakes can be very costly financially or in terms of human wellbeing.