Ming's List

View Full Calendar

Statistics Seminar - Gonzalo Mena (University of Oxford)

Event Type
Seminar/Symposium
Sponsor
Pamela Martinez
Location
1302 Everitt Lab
Date
Apr 28, 2022   3:30 pm  
Views
88
Originating Calendar
Department of Statistics Event Calendar

Title: Some statistical methods for the analysis of neural data, and pandemics.

Abstract

In modern scientific setups we are faced with unprecedented challenges regarding how to process data efficiently and in a robust way. These often reveal the brittleness of our current tools, dictating the need for new methods and motivating new questions. In this talk I will give two examples of how the pressure of formidable scientific problems necessitate the deployment of new statistical methods.

First, I will focus on my contribution to NeuroPAL, a new breakthrough technology that enables the colorful imaging of every single neuron in the brain of the C.elegans worm. I will describe new methods for two difficult tasks arising in these datasets; neural segmentation and identification, where classical methods fall short. Behind these methods there is a key statistical physics principle, the so-called Schrödinger bridge, a ‘thought experiment’ that realizes the solution of an entropy-regularized optimal transport problem. This thought experiment was proposed in 1932 but it has yet to percolate into the mainstream of statistics. I will comment on the statistical (sample complexity) benefits of entropic optimal transport and how a loss function based on this principle is a better optimization objective than the log-likelihood for clustering, reducing pathologies such as bad local optima and inconsistency. In consequence, a new algorithm derived from this loss, Sinkhorn EM, attains better, more robust neural segmentation performance. Then, I will comment on how these principles can be used for probabilistically identifying neurons in C.elegans, leading to meaningful uncertainty quantification on this hard combinatorial setup. I will further comment on how these novel methods have proven their usefulness in other contexts such as deep learning.

Second, I will present some of my work on a recent pressing problem, the analysis of the true impact of the ongoing pandemic. My work (joint with Pamela Martinez) offers a comprehensive narrative on the different pathways mediating the relation between socioeconomic status and worse outcomes.  Here, the main challenge is to eliminate the biases introduces by the reporting mechanisms. I show how the use of deconvolution models and hierarchical Bayesian models can be used to obtain more reliable understanding on the transmission dynamics and their effects on different populations. 

link for robots only