IPAB Workshop-11/06/2020

Title: ‘Just' a timely approximation Abstract: Approximation is unavoidable and forms the crux of developing practical solutions to ‘real-world’ applications. In classical approaches to  approximate modelling, the goal is to design models that mimic realistic processes as closely as possible. A more disruptive hypothesis is that the timeliness of approximation is more valuable than its deviation from reality. Developing coarse but rapid estimates enables exciting applications including robots taking decisions that require reasoning about complex dynamics (such as interacting with liquids). For such approximations to be useful, it is important to: a) develop suitable simulation methods; b) rapidly estimate parameters (mass, elasticity, etc.) from sensory data; and c) to underpin the approximation with a sound analysis of error (e.g.[4]). In this talk I will focus on my recent submission to NeurIPS 20 on approximating integrals of functions represented using neural networks. I will also briefly discuss joint work, with PhD students, towards developing an approximation for the physical simulation of brittle fracture [1] and to estimate physically meaningful parameters using robot interaction [2] and video only [3].