Omnivident Appearance-Based Gaze Estimation – Senior Project
My senior project consisted on furthering state-of-the-art work at MPII for gaze estimation. During the project, I had the pleasure of working with OpenCV, the Cambridge Face Tracker and Caffe. Further details are pending a patent application, yet my work improved accuracy results published by Zhang et al. by ~10% during simulation runs. Significant challenges included developing a background MacOS application to gather a ~100GB dataset and constructing a fully-automated pipeline that integrated a dozen custom and 3rd-party applications.
(Further info upon request)
Generating Structured Opinion-Sets from Unstructured Financial Texts
In this project for Stanford’s Deep Learning for NLP class, two buddies and I explored the use of latent word representations (CBOW, GloVe) and Long-Short Term Memories (LSTM‘s) for categorizing highly-unstructured data in financial text corpora. We used Theano for the learning, the Stanford NLP suite for pre-processing, and tapped into a Bloomberg Terminal to get our data. The “icing on top of the cake” was using a multi-armed bandit to switch representation strategies when appropriate. After the project, one of my partners went on to continue developing a company we started together; the other has started a career as a scientist for D.E. Shaw Research.
In this project, a friend and I developed a new collaborative filtering method for recommending scientific publications to interested researchers. Our project started by re-implementing Wang & Blei’s state-of-the-art system for Collaborative Topic Regression. Our improved method significantly outperformed Wang & Blei’s when performing “item cold-start” predictions and was resilient to the “user cold-start” problem. We used the Gensim package to obtain the topic modeling (Blei & Ng’s LDA) and developed a custom recommendation engine using the mrec library as a skeleton. I went on to start a company with my partner the following year.
Non-invasive Complete Blood Count (CBC)
Before switching entirely into the CS department, I, along with two friends, developed a device + method proof-of-concept for non-invasive CBC measurements. We were recognized as the most successful team in the class. The project was undertaken through Stanford’s Medical Devices Incubator, then under the tutelage of Dr. Phyllis Gardner. One of my partners became a co-founder at a dental health devices company the following year, the other started a company that develops technology for dementia caregivers.
(Further info upon request)
These are some of the most interesting and useful data science courses I took while at Stanford. They are in no particular order.
Mining Massive Data Sets
Information Retrieval & Web Search
Natural Language Processing I
Natural Language Processing II (audit)
Deep Learning for NLP
Computer Vision I
CNN’s for Visual Recognition (audit)