Confident Deep Learning

  • 3,039

Confident Deep Learning


Jinwoo Shin is currently an associate professor at the School of Electrical Engineering at KAIST, Korea. His current major research interest is on algorithmic questions for machine learning and related fields. He obtained the Ph.D. degree from Massachusetts Institute of Technology in 2010 with George M. Sprowls (Best MIT CS PhD Thesis) Award and B.S. degrees (in Math and CS) from Seoul National University in 2001. After spending two years (2010-2012) at Algorithms & Randomness Center, Georgia Institute of Technology, one year (2012-2013) at Business Analytics and Mathematical Sciences Department, IBM T. J. Watson Research, he joined KAIST EE in Fall 2013. He received Kenneth C. Sevcik Award at ACM SIGMETRICS/Performance 2009, Best Publication Award from INFORMS Applied Probability Society 2013, Best Paper Award at ACM MOBIHOC 2013, Bloomberg Scientific Research Award 2015 and ACM SIGMETRICS Rising Star Award 2015. He has served TPCs (or reviewers) at AAAI, ICC, ICML, INFOCOM, INFORMS, MOBIHOC, NIPS, SIGMETRICS, UAI, WIOPT and AEs at IEEE/ACM Transactions on Networking, ACM Modeling and Performance Evaluation of Computing Systems.



The state-of- art deep neural networks are known to be highly overconfident in their predictions, i.e., do not distinguish in- distribution (i.e., training distribution by a classifier) and out-of- distributions sufficiently different from it. To resolve the issue, we propose a new loss, called confident loss, where it forces samples from out-of- distribution less confident by the neural classifier. To show its effectiveness, we apply the proposed loss function to two classification tasks using neural networks: (a) multiple choice learning and (b) detecting out-of- distributions. This is a joint work with Kimin Lee (KAIST), Kibok Lee (Michigan) and Honglak Lee (Michigan).