Kartic Subr
Thu 20 May 2021, 13:00 - 14:00
Online (Blackboard collaborate)

If you have a question about this talk, please contact: Jodie Cameron (jcamero9)

Title: Taming Timely Inverses

Abstract: Last year, at the IPAB seminar, I spoke about the importance of timely (rapid) approximations of physical simulations. Currently, I am excited about developing rapid approximations for inverse problems involving physical simulation. Despite troves of methods for solving inverse problems, sparingly few consider timeliness as a measure of error. In many real-time applications, obtaining an accurate or precise answer is only useful if it can be estimated within a narrow temporal window. Conversely, rapid estimates are indeed less useful if they are not accompanied by analyses of popular error criteria such as accuracy and precision. In my talk this week, I will present a couple of recent works [1,2] that are early steps in this direction. In addition, I will also briefly discuss other recent work [3,4,5].

[1] Q-NET: A Network for Low-Dimensional Integrals of Neural Proxies
    Kartic Subr

[2] IV-Posterior: Inverse Value Estimation for Interpretable Policy Certificates
    Tatiana Lopez-Guevara, Michael Burke, Nicholas K. Taylor, Kartic Subr

[3] Action sequencing using visual permutations
    Michael Burke, Kartic Subr, Subramanian Ramamoorthy

[4] Jittering Samples using a kd-Tree Stratification
    Alexandros D. Keros, Divakaran Divakaran, Kartic Subr

[5] On Improved Training of CNNs for Acoustic Localisation
    Elizabeth Vargas, James Hopgood, Keith Brown, Kartic Subr