Selected Projects

Experience replay plays a crucial role in improving the sample efficiency of deep reinforcement learning agents. Recent advances in experience replay propose the use of Mixup to further improve sample efficiency via synthetic sample generation. We build upon this idea with Neighborhood Mixup Experience Replay (NMER), a modular replay buffer that interpolates transitions with their closest neighbors in normalized state-action space. NMER preserves a locally linear approximation of the transition manifold by only performing Mixup between transitions with similar state-action features. Under NMER, a given transition’s set of state-action neighbors is dynamic and episode agnostic, in turn encouraging greater policy generalizability via cross-episode interpolation. We combine our approach with recent off-policy deep reinforcement learning algorithms and evaluate on several continuous control environments. We observe that NMER improves sample efficiency by an average 87% (TD3) and 29% (SAC) over baseline replay buffers, enabling agents to effectively recombine previous experiences and learn from limited data.
L4DC 2022 Paper.

The utility of machine learning in biomedicine is being investigated in various contexts, including for diagnostic and interpretive purposes for imaging modalities, quantifying disease risk, and processing text from physician and patient reports. To best facilitate the potential of machine learning, clinicians and computational scientists must inform one another about the nature of their clinical challenges and available methods for solving them, respectively. To this end, clinicians need to critically evaluate machine learning studies conducted to solve relevant problems in medicine. This article serves as a checklist for clinicians to understand and appraise machine learning studies and help facilitate productive conversations between the clinical and data science communities to improve human health.
Intelligence-Based Medicine, 2022.

In this paper, we examine in-sample and out-of-sample classification performance of Fully Convolutional Neural Networks (FCNNs) and Support Vector Machines (SVMs) trained with and without 3D normalized digital surface model (nDSM) information. We assess classification performance using multispectral imagery from the International Society for Photogrammetry and Remote Sensing (ISPRS) 2D Semantic Labeling contest and the United States Special Operations Command (USSOCOM) Urban 3D Challenge. We find that providing RGB classifiers with additional 3D nDSM information results in little increase in in-sample classification performance, suggesting that spectral information alone may be sufficient for the given classification tasks. However, we observe that providing these RGB classifiers with additional nDSM information leads to significant gains in out-of-sample predictive performance.
Data Science Journal, 20(1), p.20.

As more autonomous systems become equipped with sensors for remote sensing, it is imperative we have methods for fast and efficient online calibration between these sensors. This is particularly important for systems with odometric drift and time-varying sensor extrinsics. In this paper, we implement a lidar-lidar calibration framework for use in the DARPA Subterranean Challenge. Using informed nonlinear on-manifold optimization to solve for an optimal relative pose in tandem with pose visualizations and both internal and external calibration scoring criteria, we develop a self-contained estimation and evaluation framework for estimating and determining the quality of lidar calibration.
16.485 Final Project

Learning competitive behaviors in multi-agent settings such as racing requires long-term reasoning about potential adversarial interactions. This paper presents Deep Latent Competition (DLC), a novel reinforcement learning algorithm that learns competitive visual control policies through self-play in imagination. The DLC agent imagines multi-agent interaction sequences in the compact latent space of a learned world model that combines a joint transition function with opponent viewpoint prediction. Imagined self-play reduces costly sample generation in the real world, while the latent representation enables planning to scale gracefully with observation dimensionality. We demonstrate the effectiveness of our algorithm in learning competitive behaviors on a novel multi-agent racing benchmark that requires planning from image observations.
CoRL 2020 Paper

Annotation of large imagery datasets remains a significant obstacle for training supervised machine learning models. Image retrieval in large, unlabeled datasets also remains a pertinent, difficult application that can be applied to image annotation and other problems in artificial intelligence. With unsupervised machine learning techniques, namely Scale-Invariant Feature Transform (SIFT), K-Means Clustering, (pre-trained) Convolutional Neural Network features, and Latent Dirichlet Allocation (LDA), this work aims to organize images in an unsupervised manner using latent features. We contend it is possible for this unsupervised image organization technique to accelerate image annotation and retrieval processes without sacrificing labeling accuracy.
6.867: Machine Learning Final Project

For our Robotics: Science and Systems course final project, my team and I implemented a system for navigating our miniaturized robotic racecar autonomously through a miniaturized urban environment. Our system was sub-divided into a vision component, which consisted of a neural-based object detection pipeline along with real-time Hough transforms and color segmentation, as well as a control system, which implemented complex driving behaviors for our robot to safely navigate in this miniaturized urban environment.
6.141: Robotics: Science and Systems Final Project

Semantic scene segmentation is a widely-used computer vision technique that is most commonly applied in a 2D RGB imagery setting. However, due to imagery's dense data structure, this computer vision task can rapidly become computationally intractable, particularly for real-time scene segmentation applications such as autonomous driving. With RGB-augmented point cloud data, however, semantic scene segmentation can be applied in a more sparse setting, enabling for more tractable training and inference for time-sensitive applications. We demonstrate baseline binary and six-class road segmentation frameworks using sparse data fusion that achieve 80% and 32% IoU, respectively. Furthermore, semantic segmentation, which is fundamentally an element-wise multi-class classification problem, suffers performance losses from unbalanced datasets. We investigate several machine learning techniques for mitigating class imbalance, namely focal loss, weighted cross-entropy, and transfer learning.
6.869: Advances in Computer Vision Final Project

Teaching Experience

Academic Courses

  • Teaching Assistant

    6.s191: Introduction to Deep Learning | Massachusetts Institute of Technology | January 2021

    Course covering the fundamentals of deep learning, including deep neural networks, Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Variational Autoencoders (VAEs), Deep Reinforcement Learning, and Generative Adversarial Neural Networks (GANs). The course is structured with both theoretical and hands-on components. More information can be found here.

  • Teaching Assistant

    6.801/6.866: Machine Vision | Massachusetts Institute of Technology | Sep - Dec 2020

    Computer vision course focused on geometric computer vision algorithms, such as edge detection, time to contact, motion estimation, camera calibration, and recovery of 3D structure from 2D imagery.

  • Laboratory Assistant

    6.036: Intro to Machine Learning | Massachusetts Institute of Technology | Sep - Dec 2018

    Introductory machine learning course that primarily covers classical and novel topics in supervised learning. Topics include SVM, regression, neural networks, CNNs, RNNs, reinforcement learning, recommendation systems, nearest neighors, and decision trees.

  • Laboratory Assistant

    6.004: Computation Structures | Massachusetts Institute of Technology | Feb - May 2018

    Introductory computer architecture course that primarily covers MOSFET structures, computational primitives such as adders and registers, pipelining, parallelism, operating systems, and Assembly.

Global Teaching Opportunities

  • Technical Instructor

    MIT Global Startup Labs | Jan 2020 | Montevideo, Uruguay

    Collaboratively developed an AWS EC2 Python module for managing cloud computing resources for our students. Developed lessons and tutorials for computer vision and Python machine learning packages, such as PyTorch, TensorFlow, and OpenCV.

  • STEM Instructor

    MIT Global Teaching Labs | Jan 2019 | Amman, Jordan

    Introduced students in Amman, Jordan to the Python programming language through collaborative exercises and interactive instruction.

Tutoring and Mentoring

  • Machine Learning and Python Tutor

    Wyzant | March 2020 - present | Remote

    Tutoring students online in machine learning, Python, artificial intelligence, and econometrics. I have over 400 hours of tutoring experience, as well as 115 5.0/5.0 ratings. You can find my profile here, though depending on the number of students I'm tutoring at the time, it may not be available. Linked below is a codebase I've created for tutoring.

  • Machine Learning Mentor

    Polygence | February 2021 - present | Remote

    As a mentor with Polygence, I advise and mentor high school students with semester-long machine learning technical projects of their choice. You can find my profile here.

  • Educator

    Numerade | April 2020 - present | Remote

    Developing videos for solutions to textbook problems in linear algebra, probability and statistics, and pre-calculus. You can find my profile here.

Leadership Experience

Project Leadership

  • Webmaster Committee Chair

    Eta Kappa Nu, Beta Theta Chapter

    Leading a project for integrating and iterating on a collaborative filtering-based class recommendation system for students at MIT. Additionally, led an effort to develop documentation for our website codebase to optimize the webmaster transition process.

  • Community Service Chair

    Tau Beta Pi, Mass Beta Chapter

    Providing community service mentorship for students to help them create meaningful impact in the local community. Past projects have included: Habitat for Humanity, Red Sox Green Team, and Girls Who Code.

  • Tech Showcase co-Director

    MIT Energy Conference

    Organized a two-hour technology showcase featuring 40 energy technology companies at the 2019 MIT Energy Conference.

  • Co-President

    MIT Undergraduate Energy Club

    Leading a team of 50 members to organize events and create opportunities in energy education and professional development. Past events have included: Working the Energy Transition forum, MIT Energy Career Fair, and energy social mixers.

  • Community Service and Philanthropy Chair

    Kappa Sigma, Gamma Pi Chapter

    Organized community service and philanthropy events for my chapter. Past projects have included: Military Heroes Campaign fundraiser, career fair t-shirt drive, and food pantry and homeless dinner service programs.