How Caltech Launched a Leading AI Conference
courtesy of Yaser Abu-Mostafa
By Kimm Fesenmaier
When Yaser Abu-Mostafa (PhD ’83) joined the Caltech faculty as an assistant professor after receiving his doctorate, AI had a bad reputation. “Nobody would say they were working in AI because it had promised much and delivered little,” says Abu-Mostafa, now a professor of electrical engineering and computer science.
That attitude was about to change, thanks largely to what was happening at Caltech. In 1982, shortly after John Hopfield, now an emeritus professor at Princeton and Caltech’s Roscoe G. Dickinson Professor of Chemistry and Biology, Emeritus, published his computer model demonstrating the basic feasibility of artificial neural networks, Abu-Mostafa organized a workshop at the Institute to discuss the model and its implications. Today, that event with just 75 attendees has grown into what is called the Conference of Neural Information Processing Systems (NeurIPS), the leading gathering of artificial intelligence researchers and professionals in the world, which attracts thousands of attendees every year.
“The Hopfield model was very inspirational for us,” says Abu-Mostafa about the recent Nobel laureate’s work. “We wanted to analyze the capabilities, see what could be done with it, and so on.” After two more years of small gatherings, the workshop had a growth spurt. In 1986, Abu-Mostafa’s former Caltech advisor Demetri Psaltis, together with Bell Labs, co-hosted a larger meeting of about 150 scientists in Santa Barbara. “That’s when we realized this is serious business,” Abu-Mostafa recalls.
In November 1987, Abu-Mostafa and Ed Posner, the late Caltech professor and JPL chief technologist, organized the first NeurIPS conference in Denver, Colorado, with sponsorship from the Institute of Electrical and Electronics Engineers Information Theory Society. Posner served as the founding general chair. As program chair, Abu-Mostafa says he “sweated bullets” knowing the conference’s success hinged on the quality of the research presented. He was so selective that he nearly rejected a paper submitted by a member of the conference organizing team and only permitted himself to present a poster about the relationship between entropy and connectivity in neural networks. “I invested a ridiculous amount of time in the program, and it paid off,” he says.
Around 600 scientists attended the first official conference, including computer scientists, biologists, mathematicians, and engineers. The first two plenary speakers were Carver Mead, Caltech’s Gordon and Betty Moore Professor of Engineering and Applied Science, Emeritus, who discussed the engineering side of “Networks for Real-Time Sensory Processing,” and Terry Sejnowski, now the Francis Crick Chair at the Salk Institute for Biological Studies and a former Caltech visiting professor, who discussed “Biological Applications of Neural Network Models.”
“If you were a scientist in this area, and you looked at the whole picture, you said, ‘This will lead to something. I don’t know what, but it’ll lead to something, and we’d better pursue it.”
“There was enough substance that if you were a scientist in this area, and you looked at the whole picture, you said, ‘This will lead to something. I don’t know what, but it’ll lead to something, and we’d better pursue it,’” Abu-Mostafa says.
The conference remains a primary forum for scientists and engineers to present and discuss developments in AI. Now organized by the NeurIPS Foundation, established by Posner, the 38th conference last December in Vancouver, British Columbia, saw more than 16,000 registrants attend along with nearly 3,000 others online.
“Every breakthrough over the last 37 years was presented at NeurIPS,” Abu-Mostafa says. These include AlexNet (2012), a model developed by a University of Toronto grad student that roundly outperformed all computer vision programs of the time; AlphaGo (2019), the Google DeepMind program that defeated human champions of the Chinese game Go; and the model for ChatGPT (2022).
“AI is no longer a bad word,” Abu-Mostafa says. “It is either an exciting word or a scary word. It is incumbent upon us, the AI researchers, to make sure that the exciting part flourishes and the scary part is eliminated or at least mitigated.”