Shea Cardozo

PhD Student in Computer Science

University of Waterloo

Hi! I’m a PhD Student in the Waterloo Intelligent Systems Lab (WISELab) at the University of Waterloo. I am broadly interested in creating generalizable autonomous agents and systems. This encompasses a wide domain of approaches such as model-based reinforcement learning techniques for planning and reasoning, causal learning and understanding, and behaviour prediction in multi-agent scenarios.

Education

PhD in Computer Science

University of Waterloo Sep 2023

Supervisor: Professor Krzysztof Czarnecki - Waterloo Intelligent Systems Lab

Master of Science in Applied Computing

University of Toronto - St. George Sep 2021 – Dec 2022

Data Science Concentration

Bachelor of Mathematics in Statistics

University of Waterloo Sep 2017 – Jun 2021

Experience

Machine Learning Scientist

Tenyks

April 2022 – August 2023 Cambridge, UK

Worked on the Research Team to come up with new data-centric ways to explain and improve Computer Vision systems.

Data Scientist - Claims AI Team

Intact Insurance

September 2020 – December 2020 Toronto, Canada

Used Vision and NLP techniques to automatically classify insurance documents.

Data Scientist - Analytics

Noom Inc.

May 2020 – August 2020 New York City, US

Improved food and support tracking to streamline internal processes.

Featured Publications

Shea Cardozo, Gabriel Islas Montero, Dmitry Kazhdan, Botty Dimanov, Maleakhi Wijaya, Mateja Jamnik, Pietro Lio

November, 2022 Presented at the AIMLAI workshop

Explainer Divergence Scores (EDS): Some Post-Hoc Explanations May be Effective for Detecting Unknown Spurious Correlations

We propose a new general methodology, Explainer Divergence Scores (EDS), to evaluate Post-Hoc Explanations for the purpose of identifying spurious correlations in neural networks. We use our methodology to compare the detection performance of three different explainers - feature attribution methods, influential examples and concept extraction, on two different image datasets.

Featured Projects

Graddy

I create is a basic ML framework implemented from scratch in Python. It comes with an implementation of AutoGrad, data structures for Tensors and Modules, basic deep learning layers such as Linear and Convolutional layers with ReLU and SoftMax activations, and optimizer classes for SGD and SGD with Momentum.

Adversarial Conditional UNET

I train a UNET model to generate adversarial examples on the CIFAR-10 dataset that can fool a variety of state-of-the-art classification models.

Painting Style Transfer

I create an Android application that captures images from the phone camera and passes the image to a GANILLA (a variant of CycleGAN) style-transfer model.