More coming soon!
My research in the robust systems lab at Northeastern University focuses primarily on activity recognition in video data using deep learning.
A survey for deep reinforcement learning in markovian cyber–physical systems: Common problems and solutions
Deep Reinforcement Learning (DRL) is increasingly applied in cyber–physical systems for automation tasks. It is important to record the developing trends in DRL’s applications to help researchers overcome common problems using common solutions. This survey investigates trends seen within two applied settings: motor control tasks, and resource allocation tasks. The common problems include intractability of the action space, or state space, as well as hurdles associated with the prohibitive cost of training systems from scratch in the real-world. Real-world training data is sparse and difficult to derive and training in real-world can damage real-world learning systems. Researchers have provided a set of common as well as unique solutions. Tackling the problem of intractability, researchers have succeeded in guiding network training with handcrafted reward functions, auxiliary learning, and by simplifying the state or action spaces before performing transfer learning to more complex systems. Many state-of-the-art algorithms reformulate problems to use multi-agent or hierarchical learning to reduce the intractability of the state or action spaces for a single agent. Common solutions to the prohibitive cost of training include using benchmarks and simulations. This requires a shared feature space common to both simulation and the real world; without that you introduce what is known as the reality gap problem. This is the first survey, to our knowledge, that studies DRL as it is applied in the real world at this scope. It is our hope that the common solutions surveyed become common practice.
This thesis paper for my Master’s degree focused on the deep learning technique called Dropout. Since 2015 and the creation of Batch Normalization researchers have advocated “dropping” Dropout from future designs claiming that Dropout and Batch Normalization used together increases the inference error. More recent research has shown this not to be the case, yet the stigma against Dropout remains. This research shows the continued utility of Dropout by simply placing Dropout layers AFTER Batch Normalization layers. I extended drop out to vary the dropout rate based on the validation accuracy – as the network learns dropout lets more and more information pass through the network unimpeded. This results in an increased accuracy during validation / testing compared to using static dropout or no dropout at all. The dynamic dropout also enforces network sparsity (like it’s static cousin).
Qualifying Exam for NEU PhD
It is official, I am a qualified PhD candidate. In my qualifying exam I briefly survey the state of the art neural network compression algorithms.
1. Action Recognition Applied in an Airport Setting
Our research is applied in the airport security domain. This is a final project for a machine learning class that explains the difficulty in applying traditional computer vision techniques in the domain of airport security. My thesis research in part attempts to overcome some of the hurdles witnessed in the course of this project.
Due to my expertise in action recognition I was asked by my professor to give a lecture alongside a peer in my advanced computer vision course.
3. Database for Historic US Election Campaigns
This is a database I designed for a final project in my database management systems class. The database tracks position statements of candidates and election results for historic US elections throughout the 1960s.
See my most recent post for a video demo!
Code available here.
For a data visualization course, my team used a regression to model the viability of proposed locations for the BlueBike company of boston which used US censor data, the google api, and publically available data from the BlueBike company of Boston.
Announcements for new locations were measured as high viability by our model suggesting our intuition matches market research.
Computer Vision & Machine Learning Coursework
The following are a collection of projects I did that are directly related to my degree.
5. Spatial & Temporal Video Filtering
6. Combining Images with Shared Landmarks
7. Circulant Tracking in Video Data
8. Extra Credit Project for EECE5639
1. Improving Music Genre Classification
This is the last project I did at Northeastern. It was for a graduate level elective in Machine Learning. This counted as the final grade in the course. I was partnered with a long time friend and colleague – Eric. We compared results of classification using text data from lyrics against using signal data from audio samples of songs of different genres. Eric studied a classifier dealing with the features extracted from audio samples of songs while I studied a classifier targeted at the songs’ lyrics. As a final step in the project we combined the two approaches we studied into a neural network that achieved a moderate increase in accuracy relative to both individual methods.
Together we got an A on the project.
This is the final project for my Senior Design project – known at Northeastern University as a Capstone Project. As seniors six of us designed and implemented an avionics board to be used as the first revision of an embedded avionics board that could be utilized in an unmanned rocket. The project serving as a proof of concept went on to be a pivotal component of the local chapter of AIAA’s attempt to be the first academic institution to breach the Karman line. In this club activity I continued on as the lead firmware developer for the project and taught many peers at Northeastern University in topics ranging from firmware development to project management. Under AIAA sponsorship several more device drivers were incorporated into the design and the communication system for the avionics system matured as well.
This is a research paper I wrote for a technical writing class while getting my degree. The paper revolves around the growing influence that biology has in the field of artificial intelligence. Many of the points being made are still relevant today!