Autonomous vehicles (eg. self-driving cars) present huge potential to improve quality of life and efficiency of transportation. Recent studies leveraging deep learning methods have demonstrated it is possible to learn internal representations of end-to-end processing steps for controlling steering from vision data. Traditional methods tackle autonomous steering through explicit decomposition, leveraging vision data to detect road features, such as lane markings and outline of the road, followed by path planning and control based on the outputs of the visual detection algorithms. Conversely, the end-to-end deep learning approach takes pixel inputs of the scene and augments intermediate models to directly output a control command. However, this potential is extremely powerful, training such a large deep neural network is computationally intensive, and the corresponding equipment used is expensive. Additionally, in many scenario's these techniques may not be feasible due to changing environmental conditions or even lack of resources. In this project we aim achieve additional robustness through incremental, distributed training of the end-to-end steering control algorithm leveraging low powered, low-cost embedded devices that could be installed on-board autonomous vehicles.
This project proposes to push the limits of autonomous vehicles in terms of speed, precision handling, and safety. We modeled and simulated the dynamical vehicle system in Drake, while using system identification capabilities based on real-data collected from on-board sensors and Drake simulator. Compared to existing systems, which are based solely on physics based models, our use of models derived from real world driving conditions and a physical vehicle provide a more realistic simulation and novel insights into the complex variety of variables encountered in a real driving scenario.
In this project I evaluated the effectiveness of Stochastic Variance Reduced Gradient descent in Deep Neural Networks. Despite being widely tested in the convex case, many of these optimization through variance reduction techniques have not been accurately and fairly compared in the non-convex case, and in specific Deep Learning networks. I created, trained, and tested Convolutional NN (CNN), Recursive NN (RNN), Deep Belief Networks (DBN), and implemented, integrated and evaluated relevant optimization algorithms (SVGD, AdaGrad, AdaDelta, etc) in Theano, and assessed against MNIST and IMDB datasets.
Research internship in MIT's SENSEable City lab with Principal Investigator, Professor Carlo Ratti, and graduate and post-graduate researchers. I am utilizing data mining and machine learning techniques to analyze mobile phone data in order to understand the differences in how people move and interact in high risk, developing regions such as the Ivory Coast, versus those in more stable and developed regions, such as Europe and the US. The goal is to illuminate the vulnerabilities of people in such regions, with a potential result of providing governments and aid organizations with tools to better address societal needs through urban planning, infrastructure, and services such as public transportation, mobile banking, social programs.
I am developing mathematical models in Matlab, and creating software tools for geospatial and temporal visualizations of the data and models. My project is part of a grassroots initiative called Developers for Development, in which software developers and data scientists volunteer their expertise to help developing nations.
Research internship in the Contextual Content Analysis lab with Principal Investigator, Professor Noel O'Connor, and graduate and post-graduate researchers. During this internship I developed a real-time system for tennis sensor data analysis and feedback for motion refinement, based on algorithms I had previously discovered. The system I designed and developed includes an offline training component, which uses advanced machine learning algorithms to train player-specific models for use by the real-time motion refinement system. The real-time system captures, splits, translates, calibrates, de-noises, normalizes, and computes features from the wireless inertial sensor data; it detects potential strokes, identifies the type of stroke and deviations from reference strokes; and finally, it generates precise, and unbiased feedback to the player or coach - all in real-time.
Collected 400 GB of tennis sensor data from a combination of wireless inertial sensors, video cameras, microphones, and image cameras. The elite Sanchez-Casal Tennis Academy is world-famous for training tennis stars such as Rafael Nadal and Andy Murray. I proposed and was accepted to collect sensor data from 24 players ranging from beginners to ranked players over the course of 1 week. I prepared for the trip by evaluating a variety of sensors, designing a data collection methodology, and developing software tools to capture, organize, and validate data as it was being captured. Additionally, because the volume of data captured was so large, I also had to subsequently design and develop software tools to cleanse and annotate the data so it could be used to evaluate my feature extraction, model building, and model test algorithms.
Top prizewinner in the 23rd European Union Young Scientist competition for project: Tennis Sensor Data Analysis: An Automated System for Macro-motion Refinement. The competition was restricted to 87 student winners of their respective national science competition, aged 21 and under, from 37 countries. Projects are assessed on scientific achievement and jury interviews. The 3 top prizewinners each receive €7000, and trips to the London International Youth Science Forum and Nobel Prize ceremonies in Stockholm, Sweden.
Key results of my project included an analysis of over 400 GB of tennis sensor data, and development of models and algorithms for detecting the subtle differences in motion and orientation that enabled automatically distinguishing between 15 different tennis stroke types (e.g., a forehand flat versus forehand topspin) using only inertial data. I was able to achieve 96% accuracy across all stroke types, with 98% accuracy for 11 of the 15 strokes, and 99% or higher for 6 of the 15 strokes.
My results were tested rigorously through cross validation over millions of samples and on a variety of skill levels, physical builds, and geographies. I also designed and developed a real-time computer system to demonstrate the accuracy and low latency of my algorithms, as well as the capability to provide the player or coach with immediate, precise, and unbiased feedback on how to refine their motions and orientations for improved performance.
My results are significant because previously published research relied on video and inertial sensing, which requires a fixed and significantly more expensive infrastructure, and focused on identifying stroke classes (e.g., forehand versus backhand), which is too coarse-grained to support computer-assisted skill refinement.
Awarded scholarship for funded summer internship at the CLARITY Centre for Sensor Web Technologies at Dublin City University. Established in 2008 with €16M funding, CLARITY is 1 of 10 national Centers for Science, Engineering and Technology (CSET) in Ireland. CLARITY's graduate and post-graduate researchers investigate next generation, adaptive sensor technologies, and the analytic algorithms required make sense of the generated data.
Grand prizewinner in Ireland's national science competition for project: Tennis Sensor Data Analysis. In 2011, there were over 550 entrants in areas including biology, chemistry, mathematics, physics, social sciences, and technology, and are assessed on scientific merit of their technical report, poster, and jury interviews. The grand prize is €5000, a Waterford crystal trophy, and the opportunity to represent Ireland at the European Union Contest for Young Scientists 2011. http://www.btyoungscientist.ie/media/pressreleases.php?id=60
Starting member of Yorktown Heights High School Senior Boys Tennis team in my freshman year. I was a winner at the League and Division Championship level, and was awarded the chance to compete in the New York State Sectional competitions.
Completed courses include (click to expand):
Graduation: June 2012
Honors Curriculum, Math and Sciences concentration
Freshman: 2009-2010, GPA: 3.8/4.0