Daniel Herrera-Esposito

Daniel Herrera-Esposito

Computational Neuroscientist. Interested in the brain, math, computers, and sometimes epidemiology

Talks and presentations

Supervised Quadratic Feature Analysis: An information geometry approach to dimensionality reduction

December 14, 2024

Poster, Symmetry and Geometry in Neural Representations Workshop, NEURIPS 2024, Vancouver, Canada

Supervised dimensionality reduction seeks to find a low-dimensional feature space that maximizes the separation between classes. We propose a novel supervised dimensionality reduction method, Supervised Quadratic Feature Analysis (SQFA), that exploits the geometry of the data second-order statistics to maximize second-order class separation. Poster.

Noise and divisive normalization: Neuroscience in online math forums

November 07, 2024

Invited talk, Society of Neuroscience of Uruguay (SNU) meeting., Montevideo, Uruguay

Neural noise is an ubiquitous phenomenon in neural systems. Better understanding its origins and consequences is crucial for understanding neural coding. In this talk I explore the relation between neural noise and divisive normalization. I also discuss my use of online math forums to develop the mathematics behind this modeling work. Slides

A geometric analysis of task-specific natural image statistics

August 02, 2024

Invited talk, Alex H. Williams lab meeting, Flatiron Institute, New York, NY, USA

There is growing interest in the geometric analyses of representations in biological and artificial neural systems. Generally, these analyses do not consider the stochastic nature of the representations. Here, we use differential geometry to analyze the geometry of response statistics in a simple ideal observer model to naturalistic images across different visual tasks. We find that different information geometric structures are required for different analysis goals. Slides

Image-computable Bayesian model for 3D motion estimation with natural stimuli explains human biases

December 14, 2022

Poster, Shared Visual Representations in Humans and Machines (SVRHM) Workshop @ NEURIPS 2022, New Orleans, LA, USA

How humans use different binocular cues to perceive 3D motion is still not well understood. In this work we develop an image-computable Bayesian model for 3D motion estimation and train it on naturalistic binocular videos. The model shows behaviors similar to those reported in human psychophysics. Poster,

Age-stratified severe and critical infection rates to estimate COVID-19 under-ascertainment

July 20, 2021

Talk, 7th International Conference on Time Series and Forecasting, Virtual

Underascertainment of SARS-CoV-2 infections is a common problem during the COVID-19 pandemic. Estimating the degree of underascertainment can be challenging, but it is useful for guiding public health responses. Here we estimate underascertainment across time for Uruguay during 2020, adapting a method developed by Russell et al. (2020) to incorporate age-stratified case reports, as well as for using hospitalization and ICU admission data. Video

Crowding of naturalistic visual textures

March 01, 2017

Talk, Uruguayan Society of Biology Meeting 2017, Montevideo, Uruguay

Summary statistics representations of visual inputs are used to model visual texture perception. These type of representations are also used to model visual crowding in peripheral vision, in the so called texture-tiling model of peripheral vision. Here we probe visual crowding using naturalistic textures, to test how well the texture-tiling model can account for perceptual organization in crowding. Slides