Feb 16, 2024 : CSE Bytes
Prof. Harish Guruprasad
Title: Feature learning in neural networks
Brief Abstract: "What makes neural networks work better than classic methods like linear/kernel algorithms?" is still a wide-open question. The candidate answers have ranged from "neural networks implement kernel methods efficiently" to "neural networks can hierarchically identify disparate regions of the input space that have similar label structure". In this talk I will go over some of these candidate answers, give arguments for they might not be completely correct, and also give a new hypothesis based on discontinuities in the label-function and interacting learnable features.
Prof. Krishna Nandivada
Title: Myths, and Facts about performance. Or how to realize performant codes.
Brief Abstract: We will try to understand two popular myths: (i) Parallel programs are always faster than their serial counter-parts. (ii) Applications written in languages like Java always run slower than their counterparts in languages like C/C++. We will try to see how we are trying to establish one of these two myths and bust the other one.
Date: Friday Feb 16
Time: 3.30 - 4.30 PM
Venue: SSB 134
