I’m currently an AI Scientist at Exo, an amazing company that aims to change the landscape of Point-of-Care Ultrasound imaging. I’m part of the AI team, whose objective is the development of smart algorithms that help clinicians to make better and faster decisions when interpreting medical ultrasound images.
I did my Ph.D. in Statistical Machine Learning at the University of Alberta, working under the supervision of Profr. Russ Greiner. I was also part of the Alberta Machine Intelligence Institute and a former member of the Computational Psychiatry Group. My thesis was about the use of machine learning in medical applications when we have small training datasets (Watch the presentation here).
My current research focuses on finding learning methodologies that are sample efficient, allow the incorporation of domain expert knowledge, produce accurate and calibrated models, and produce classifiers that are robust enough to handle images with different properties. I’m approaching this goal from two different perspectives: (1) the incorporation of additional information, based on medical expert knowledge, into the learning process, and (2) techniques from domain adaptation to improve the robustness of models.
We recently proposed a methodology for learning classifiers using probabilistic labels, instead of the traditional categorical labels. The main idea is to replace the categorical labels (usually encoded as one-hot vectors) with probability mass functions. These new labels give a learner more information per training instance, allowing it to learn useful concepts with fewer instances. Additionally, they lead to models that are calibrated, which give a straightforward interpretation of the output as the probability of diagnosis. To learn such probability mass functions we proposed to use probabilistic graphical models that encode information that is relevant for classification. This was published at AISTATS’21 (See the paper here, and a talk on the topic here). Similarly, we are currently exploring approaches based on multitask learning and self-supervised learning for learning useful representations with limited training instances.
Another example is the forecasting of people infected with COVID-19. In this paper we explain how we combine machine learning with probabilistic graphical models and epidemiological models to make accurate predictions. You can also watch the talk here.
Because of not-well-understood reasons, data coming from different scanning sites have different probability distributions, even if they are observing the same phenomena! This complicates the job of machine learning algorithms, and it is a problem known as ‘Batch effects’ (A problem closely related to Domain Adaptation and Transfer Learning). Since the assumption is that all sites are observing the same phenomena, there must be a common representation that is site-independent. How to find this representation is the focus of my research. One hint: Z-score normalization or whitening of the data is not enough (to know why you can read my M.Sc. thesis).
A second area of interest is the development of undirected graphical models that combine continuous and discrete data. These models become relevant in medical diagnosis for several reasons. For example, they allow us to visualize the statistical dependence between the variables, and provide a framework to work with missing data, which is common in medical applications. These models are a good alternative to working with data coming from different sources: imaging, genetics, surveys, etc.
Academic activities
Teaching assistant
– Probabilistic graphical models (Winter 2016, Winter 2017, Fall 2018, Fall 2020)
– Introduction to machine learning (Fall 2015, Fall 2016, Fall 2017, Winter 2019)
– Introduction to the foundations of computation II (Fall 2014, Winter 2015)
Relevant courses
– Introduction to machine learning
– Probabilistic graphical models
– Natural language processing
– Topics on deep learning
– Techniques of mathematics for statistics
– Stochastic processes
– Optimal transport
Education
Ph.D. Statistical Machine Learning (Jan 2017 – Sep 2022)
University of Alberta – Edmonton, AB, Canada
Supervisor: Russ Greiner
Thesis: Machine learning for medical applications with limited data: Incorporating domain expertise and addressig domain-shift [pdf](Nominated for Best Thesis Award at University of Alberta, Computing Science Department)
M.Sc. Computing Science (Sept 2014 – Dec 2016)
University of Alberta – Edmonton, AB, Canada
Supervisor: Russ Greiner
Thesis: The challenge of applying machine learning techniques to diagnose schizophrenia using multi-site fMRI data. [pdf] (Nominated for Best Thesis Award at University of Alberta, Computing Science Department)
B.Sc. Mechatronics Engineering
Tecnológico de Monterrey (ITESM) – Guadalajara, Jalisco, México