Blake Martin

I am a Machine Learning Master's student at Carnegie Mellon University. My current research is focused on multilingual machine translation with the MultiComp Lab.

I completed my undergraduate studies in Data Science at the University of Michigan where I worked with Professors Sindhu Kutty and Mithun Chakraborty to design and study probabilistic belief aggregation mechanisms. Previously, I also applied data-driven optimization methods to develop fetal brain models with the Computational Physics Group. I have additionally worked on computer vision and optimization projects at PathAI, KLA and Amazon.

My interests are broadly in using machine learning and computer vision to drive impactful solutions. I am particularly interested in human-centered applications including health and behavior.

Email  |  Resume  |  LinkedIn

Under sophisticated trader models, we study a particular incentive-compatible prediction market design that maintains an aggregate beta belief distribution over a bounded, continuous random variable as is the case in modeling election vote-share. We design a Bayesian trader model in which each trader uses the aggregate of information from past trades to form their prior beliefs, and then updates their posterior beliefs based on private observations. Our results suggest that early arrival can significantly dominate both wealth and informativeness in determining an agent's compensation.

A data-driven approach to inferring growth and deformation during early development of the fetal human brain. Using the fetal brain atlas MRI data, we combine direct and adjoint optimization methods to produce a model that obeys the laws of morphoelastic growth with controllable error.

An exploration of the interplay between trader informativeness, budget, and sequence in a new aggregation mechanism: the Beta-Bernoulli Prediction Market. Our results indicate desirable aggregation properties and suggest that early arrival can significantly dominate both wealth and informativeness in determining trader compensation.

An introductory study displaying the power of machine learning applied to initial screening applications in cytogenomics. We show that convolutional neural networks can be trained to accurately identify abnormal chromosomes associated with chronic myeloid leukemia.

I lead office hours and a discussion section teaching concepts including SVMs, neural networks, decision trees, boosting, and clustering. I also explore solutions to accessibility challenges posed by the virtual semester.

I improved the accuracy and robustness of tissue segmentation models with recent advancements in vision-based attention mechansims, including the neighborhood attention transformer.

I implemented heuristic solutions of the vehicle routing problem to optimize the efficiency of driver transport tours.

I explored different self-supervised representation learning techniques to capture relevant information in high resolution diffraction pattern images. My work with autoencoders and simCLR enabled highly predictive encodings to be produced three times faster than with previous methods for transfer learning tasks.

I extracted accurate predictions of thermal sensation and comfort from 2D car seat pressure heatmaps and test drive questionnaires. My work applying machine learning models and ensemble methods enabled sensor area needed for accurate predictions to be reduced by 98%. The company is currently filing a patent based on the results of my project.

I initiated this effort to provide seniors age 50+ access to guidance with technology. I volunteered as both a one-on-one instructor for personalized help as well as a presenter to classes of 10-20 seniors.

Adapted from this template