1. Intro
My Academic Research
Augmented Reality Tracking in Air Traffic Control (ATC) Operations
Jason D. Lazar | June 2024
Project Mentor: Tianyuan (Roger) Dai
CS231A: Computer Vision, From 3D Perception to 3D Reconstruction and Beyond
It is well established that the aviation industry relies on a robust air traffic control (ATC) network to maintain safe, redundant, and effective operations both on the ground and in the air. However, one of the greatest barriers to safety at major airports is extreme, low-visibility weather conditions. Thus, it is imperative to begin implementing and integrating systems to increase aviation safety when operating in suboptimal conditions. To do so, this paper introduces a relatively simple, yet effective computer vision approach to implement a real-time virtual display that will help aircraft controllers identify aircraft from the control tower, in addition to their radar screens. By using radar data provided by Flight Radar 24, I manually extracted top-down coordinates with respect to a bounded region. Subsequently, I manually defined correspondences from a static alternate perspective of the aircraft, and computed an optimal perspective homography H*, allowing the system to effectively track an aircraft in a toy example. Finally, I conducted the same methodology on dynamic footage in low visibility scenes, computing a homography matrix for each frame H1, ..., Hn, and transforming each radar coordinate. Since the system is built on ground truth radar data, I achieved 100% tracking accuracy for both static and dynamic cameras on real-world footage. Compared to openCV’s mean point algorithm for object tracking and CNN-based motion detection, my system worked incredibly well while the other systems did not work when introduced to Gaussian noise and occlusions.
Increasing the Efficiency of the Sophia Optimizer
Jason D. Lazar & Caia M. Costello | March 2024
Project Mentor: Kaylee C. Burns
CS224N: Natural Language Processing with Deep Learning
Pre-training large language models (LLMs) is incredibly time-consuming and costly. Our project aimed to improve the efficiency of Sophia, a second-order optimizer proposed by Stanford researchers that achieves 2x convergence efficiency over Adam. To do this, we propose a new way to estimate the Hessian diagonal that Sophia requires, by collecting the average of the square of the gradients directly. This approach is referenced in the original research, but it was not implemented. This significantly improves Sophia, offering an improved optimizer for language model pre-training. On a 30M GPT2 model, our optimizer (CAIA Continuous Adaptive Information Averaging) converges more than twice as fast as Sophia and more than four times faster than AdamW.
Nominated & Awarded as an Outstanding Project
Recruiting in the White House: Addressing Fairness in Automated Hiring Algorithms
Jason D. Lazar, Proud B. Mpala, Eva Prakash, Josh Singh
Project Mentor: Charlotte Peale
CS256: Algorithmic Fairness
It is evident that the job application process is only becoming more competitive for both entry-level careers and experienced professional roles. As applications continue to increase, automation is becoming an integral part of the recruitment process. It is well established that automated systems are currently in place to auto-screen applicant resumes and listed qualifications. Systems like Applicant Tracking Software (ATS) are implemented to save recruiters time and effort as they work to narrow down potential candidates. However, we predict that as artificial intelligence (AI) and machine learning models are increasing in capability, such systems will begin to automate HR-related components to job applications, like initial behavioral interviews. More specifically, we envision a future where companies will rely on trained models to conduct virtual interviews by using both computer vision and NLP techniques. It is likely that more decision making power will be entrusted to artificial models, saving companies time, effort, and money. Thus, when selecting candidates, it is imperative to address fairness concerns to ensure an equitable and impartial hiring process, as most companies promise to the public. This paper focuses on investigating the fairness implications of using automated hiring algorithms in the context of White House recruitment. By analyzing the application of AI to conduct interviews, we aim to address individual and group fairness concerns. Through a mathematical framework, we asses the extent to which these algorithms adhere to fairness principles and propose strategies to mitigate biases and promote equitable hiring outcomes. This research contributes to the broader conversation on algorithmic fairness in recruitment processes, with implications for improving diversity and representation in political institutions
Kalman Filtering For Heading and Altitude Prediction Under Uncertainty
Jason D. Lazar
Project Mentor: Dylan Asmar
AA228: Decision Making Under Uncertainty
Sensor readings are uncertain and noisy, especially for small aircraft like the Cessna 172. There are many instances where key systems can fail, and provide inaccurate readings, especially in harsh conditions. To increase safety in aircraft operations, the question becomes: how can we create a redundant, robust backup system for small aircraft, providing the pilot the ability to maintain heading, pitch, and altitude while troubleshooting errors? My research leverages the mathematical technique of Kalman Filtering to address this problem and provides a starting point for the future development of a robust backup system for a small aircraft in uncertainty. Using previous research and MATLAB's Simulink, I used model-based systems engineering (MBSE) to simulate a 3DOF aircraft model to test and validate this system.