ANC Seminar - Antonio Vergari and Magdalena Navarro
Tuesday, 26th October 2021
"An Ode to Tractable Probabilistic Modeling" + "Automating Probabilistic Reasoning" - Antonio Vergari
Abstract: In this talk I will first motivate why you should care about tractable probabilistic inference routines and show some of the perks of having them by providing examples taken from my past and current research. Then, I will focus in the second part on how the quest for tractability can be (partially) automated so that we can distil efficient and reliable inference algorithms in a systematic way for computing many of the complex queries we commonly use in ML and AI that involve integral expressions over generative and discriminative models. E.g., entropies, divergences, moments, etc.. This is part of a work that just got accepted as oral at NeurIPS2021.
Correcting label bias in classification methods to infer new putative genes involved in Autism Spectrum Disorder - Magdalena Navarro
Abstract: Training classification algorithms using datasets that contain biases in their labels can result in the classifiers inheriting these biases and unfairly disadvantaging certain groups. This issue, in conjunction with the growing adoption of Machine Learning algorithms in real-world applications, with a direct impact on people's lives, have made fairness in classification an increasingly important concern and have led to the development of several bias correcting algorithms. In this talk I will present a modification to a bias correction algorithm proposed by Jiang and Nachum* that allows for correction of continuous (instead of binary) bias in the labels and demonstrate its use in correcting the bias in a classification model for the inference of new putative genes involved in Autism Spectrum Disorder. * Jiang H, Nachum O. Identifying and Correcting Label Bias in Machine Learning; 2019.
Event type: Seminar
Date: Tuesday, 26th October
Speaker(s): Antonio Vergari, Magdalena Navarro
Chair/Host: Svitlana Briachenko