Matthew Allen

In 2020, I earned a Bachelor of Science degree in Computer Engineering Technology from the Rochester Institute of Technology. During my studies, I immersed myself in machine learning, teaching myself everything I could on the subject.

My dedication led to an internship at the Center for Systems Biology, a multidisciplinary biotechnology research lab at Massachusetts General Hospital. There, I developed novel image processing techniques based on machine learning to identify potentially cancerous cells in tissue sample images taken with digital holography.

Since then, I have pursued research in reinforcement learning at the same lab. My work has involved fundamental questions about the nature of exploration in decision processes. In this pursuit my co-author and I developed a novel mathematical framework for agent behavior and a highly scalable finite differences algorithm for deep reinforcement learning. Currently, I am investigating improvements to modern off-policy learning algorithms.

Skills

I am a programmer with experience in a variety of languages including C, C++, Python, Java, and others.

• Machine Learning - Experienced in designing and developing sophisticated and scalable machine learning software systems with PyTorch.
• Tools - Proficient with tools for highly scalable systems such as Redis and gRPC.
• Version Control - Familiar with workflows involving Git, SVN, and other version control services.
• Team Collaboration - Skilled in teamwork and collaboration, having contributed to several team-oriented research projects.
• Hardware Development - Trained in analog circuit analysis and hardware development with VHDL.
• Open Source - Successfully managed and contributed to thriving open-source projects.

Accomplishments

I've spent my life working with computers and code. Here are a few of my proudest accomplishments.

Real-time Low-cost Lymphoma & Breast Cancer Detection

I designed, developed, and tested deep learning image processing methods to automatically detect and classify potentially cancerous cells in real-time from a digital hologram. Combined with a GUI I designed for technicians to interact with my technology, I deployed this system on a Raspberry Pi. This is currently being used in Botswana to provide low-cost cancer screenings for people who may otherwise be unable to afford them. The development of this technology led to two publications in Nature and ACS Nano.

Rocket League Gym

My passion for gaming and reinforcement learning led me to co-author a mod and API for the popular Esport Rocket League, enabling ML practitioners to control the game via Python as though it were a Gymnasium-style learning environment. We titled this project Rocket League Gym (RLGym), and it has gained a large following of bot-makers. Users of our project developed a high-level open-source bot called Nexto with reinforcement learning. Trained with crowd-sourced compute power, Nexto became so proficient that it caught the eye of popular YouTube content creators like Sunless Khan, Lethamyr, and Rocket Sledge. Across all videos about Nexto, the RLGym community bot has accumulated over 1 million total views.

A Scalable Finite Difference Method for Deep Reinforcement Learning

In my first author work, I developed a highly scalable distributed system to train reinforcement learning agents with a novel finite differences algorithm. The platform architecture uses gRPC to facilitate communication between a central learning process and numerous connected remote worker processes. The system is fully deterministic and integrated with Weights & Biases for hyperparameter sweeps and repeatable experiments. My paper on this project can be found here, and starter code for the scalable finite differences algorithm and an implementation of the generalized Novelty Search algorithm from my prior research can be found on my GitHub here.

Contact Me

Email: matthew.allen.william@gmail.com
LinkedIn
GitHub