November 3, 2016
Deep Learning: the promise and the pitfalls
California Institute of Technology
Powell-Booth 100 (Seminar Room)
Abstract: Deep Learning has rapidly transformed neural networks from a black box to a black art. For certain set of problems its performance far exceeds any other technique we have seen. The issue is nailing what the 'it' is. If the GoogLeNet, ResNet etc. are an indication, just like in the winning entry of the Netflix challenge, we are now seeing a wide variety thrown into the mix when concocting powerful deep networks. The development of DeConvNets and an ability to create adversarial examples to train networks better has helped with improving their performance immensely. We look at some of the basics behind different aspects of the Deep Networks, and identify a few problems where we plan to apply them.
Bio: Ashish Mahabal is a Sr. Research Scientist interested in Data and Data Science. Trained as an astronomer, his focus is on real-time classification of transients - objects that change in brightness over short amounts of time-scales. he has been part of many astronomical sky surveys, and is involved in Earthcube, and the Early Detection Research Network (EDRN) for cancer, and in general in methodology transfer.