Matthew Kehoe

Matthew Kehoe

Data and research scientist with a background in computational mathematics. Passionate about machine learning and scientific software development.

Talks and presentations

Physics-Informed Neural Networks (PINNs) and Scientific Machine Learning

February 06, 2025

Talk, Ann Arbor Machine Learning Group, Ann Arbor, MI

Physics-Informed Neural Networks (PINNs) have transformed the numerical solution of partial differential equations (PDEs) by embedding them as soft constraints within the neural network training process. This approach has made PINNs a key component of the scientific machine learning (SciML) ecosystem. In this talk, we provide a broad overview of the PINN architecture, followed by three examples from physical systems: a damped oscillator with sinusoidal forcing, the Helmholtz equation, and a second-order elliptic eigenvalue problem known as the Gross-Pitaevskii equation. Furthermore, we demonstrate how PINNs can be used in practice and discuss key challenges, such as automatic weight selection via self-adaptive weight balancing, selecting the appropriate optimizer, and irregular boundary conditions. Notebook.

Tokenization, Sequences, and Attention Masks with Hugging Face

September 24, 2024

Talk, Applied Machine Learning Group, Ann Arbor, MI

Covered the fundamental concepts of Transformer models, including tokenization and attention mechanisms. Learned how to use tokenizers to convert text into model-friendly tensors and explored the interaction between tokenizers and models for generating predictions. Reviewed important aspects such as input IDs and attention masks, building a foundation for navigating the Hugging Face documentation and implementing these models in machine learning tasks.

Transformers and pretrained models with Hugging Face

June 06, 2024

Talk, Applied Machine Learning Group, Ann Arbor, MI

Explored the fundamentals of transformers and their application through pretrained models using Hugging Face. Reviewed key concepts and practical methodologies, providing a foundation for integrating these models into various machine learning workflows.

Scaling-up model training with GPUs and TPUs

May 14, 2024

Talk, Data Science and Machine Learning Collaborative Learning Group, Online

Explored key methods to boost model performance, such as hyperparameter tuning and model ensembling. Additionally, we reviewed ways to speed up and scale training using multi-GPU and TPU setups, mixed precision, and cloud computing resources.

Generative Adversarial Networks and Unsupervised Learning

February 13, 2024

Talk, Data Science and Machine Learning Collaborative Learning Group, Online

This talk summarized different aspects and intricacies of generative adversarial networks and a unsupervised machine project for classifying mall customers by clustering.

Neural Style Transfer, Variational Autoencoders, and Supervised Learning

January 09, 2024

Talk, Data Science and Machine Learning Collaborative Learning Group, Online

This talk summarized neural style transfer and variational autoencoders for image generation, image augmentation, and image blending. In addition, we also reviewed a supervised learning project focused on predicting future land and ocean temperatures using multiple regression models with US temperature data.

Transformers and Natural Language Processing

September 12, 2023

Talk, Data Science and Machine Learning Collaborative Learning Group, Online

This talk described the encoder-decoder architecture for Transformers and self-attention as a weighted combination of all word embeddings. In addition, several advantages were presented over RNNs and ConvNets.

Deep Learning for Timeseries

July 11, 2023

Talk, Data Science and Machine Learning Collaborative Learning Group, Online

This talk discussed time series forecasting, automatic learning of temporal dependence, and how neural networks are able to automatically learn arbitrary complex mappings from inputs to outputs and support multiple inputs and outputs.

Interpreting what convnets learn

May 09, 2023

Talk, Data Science and Machine Learning Collaborative Learning Group, Online

A fundamental problem when building a computer vision application is that of interpretability. This talk discussed various ways machine learning models can make results easier to interpret by humans. Slides.

Calculating Zeros of the Riemann Zeta Function

November 22, 2021

Talk, UIC Math Club, Chicago, Illinois

Summary of the Euler-Maclaurin Summation Formula, Riemann–Siegel Formula, and Odlyzko–Schonhage Algorithm in relation to calculating zeros of the Riemann zeta function. Slides.

High-Order Pertubation of Surfaces (HOPS) Methods

February 11, 2021

Talk, UIC Graduate Analysis and Applied Mathematics Seminar, Chicago, Illinois

Overview of the High-Order Pertubation of Surfaces (HOPS) schemes and their relation to Spectral Element Methods.

The FROST and FROSTb Models

August 12, 2020

Talk, Cold Regions Research and Engineering Labatory (CRREL), Hanover, New Hampshire

Summary of work done as a summer intern at CRREL. Slides.