Research
Various projects that I’ve taken on.
Visual Reasoning
I worked to develop a visual reasoning system that performs Visual Question Answering, evaluated on the CLEVR dataset. The resulting framework, called Transparency by Design (TbD), achieved state-of-the-art 99.1% accuracy on CLEVR and state-of-the-art accuracy on the CLEVR-CoGenT dataset.
Built around a visual attention mechanism, this model is transparent and straightforwardly interpretably. By visualizing the model attention at each operation in answering a question, we can gain insight into the TbD model, diagnose issues as they arise, and easily develop improvements to the model architecture. For more information, read the paper or get the code! You can also view the poster!
Clustering by Compression
Utilizing conditional complexity from algorithmic information theory to cluster and classify galaxies. For more information, see this page at the Iowa Space Grant Consortium.
Semantic Segmentation
I spent a summer at MIT Lincoln Laboratory working on improving semantic segmentation systems on thin objects and very large objects, while building functionality needed for segmentation into a deep learning framework based on Theano. I modified the SegNet model and ended up developing an architecture very similar to U-Net.
MIT Mahali Project
I worked at MIT Haystack Observatory on the Mahali Project, which aims to use smartphones for crowdsourcing atmospheric science data collection. I developed a scientific data processing framework for Android called MCheetah, then used this framework for a Mahali proof-of-concept and for collecting magnetometer data. Get the code!
Indoor Localization
Using machine learning and the embedded sensors in mobile devices to achieve indoor localization errors as low as 0.76 meters on average. Read the paper or get the code.
Automated NMR Assignment
Automating the assignment of backbone protein sequences using data form nuclear magnetic resonance spectroscopy experiments. Read the paper or get the code