Despite recent advances in deep learning, we do not know yet how we can combine these application-specific models to build an artificial general intelligence (AGI). Furthermore, the data is becoming the bottleneck to scale these approaches for the multiple tasks. In this talk, I propose a theory of the thinking and a neural algorithm that can bootstrap intelligence with limited computational resources and data. This neural algorithm approximates the O(n3) parameter space of the thinking theory into the O(1) parameters to make learning tractable for the biological intelligent agents. I will explain this proposal by cognitive phenomenon that are observed in a human, such as infant language acquisition, visual and verbal thinking, personality, creativity, exploit-exploration trade off, dreaming, one-shot learning, abstract language.

Deokgun Park earned his doctoral degree from the University of Maryland in 2018. His research interests include open-ended tasks, visual analytics and text mining. He leads the Human Data Interaction Lab, which studies the way humans conduct exploratory data analysis, interactive tools for text mining, and testbed for artificial general intelligence (AGI). He earned honorable mention best paper honors at the 2016 ACM Conference on Human Factors in Computing Systems and organized the IEEE VIS 2017 VADL Workshop. He has licensed patents to companies including Samsung Electronics. He completed M.S. in Interdisciplinary Engineering at Purdue University and M.S. in Biomedical Engineering at Seoul National University, where he obtained B.S. degree in Electrical Engineering. He worked at the Government Research Institute, Industry Research Labs, and startups.