Cell Community Biophysics

Individual Project, Biophysics, Simulation

(I will activate above link once my first-author work is published.) Confluent populations of elongated cells give rise to ordered patterns seen in nematic phase liquid crystals. I correlate cell shape and intercellular distance with intercellular alignment using an amorphous Ising-like model. I compare in vitro time-lapse imaging with Monte Carlo simulation results by framing a hard ellipses model using Boltzmann statistics. Furthermore, I find a statistically distinct alignment energy at quasistatic equilibrium among fibroblasts, smooth muscle cells, and pluripotent cell populations when cultured in vitro. These findings have important implications in both non-invasive clinical screening of the stem cell differentiation process and in understanding cell monolayer biomechanics from a macro-perspective.

Relating Critical Ising Models With Convolutional Neural Nets

Individual Project, Neural Networks, Ising Model, Renormalization Group

In this project, I related statistical mechanics and deep learning by replicating a result of a paper I read. The abstract is as follows: Convolutional neural net-like structures arise from training an unstructured deep belief network (DBN) using structured simulation data of 2-D Ising Models at criticality. The convolutional structure arises not just because such a structure is optimal for the task, but also because the belief network automatically engages in block renormalization procedures to "rescale" or "encode" the input, a fundamental approach in statistical mechanics. This work primarily reviews the work of Mehta et al., the group that first made the discovery that such a phenomenon occurs, and replicates their results training a DBN on Ising models, confirming that weights in the DBN become spatially concentrated during training on critical Ising samples.

Differential Convolutional Networks for Video Classification

Individual Project, Deep Learning, Computer Vision, Biology

I studied differential convolutional neural networks (dCNNs) for timelapse video classification of cell mixtures. The purpose of the differential convolutional neural network is to study the timelapse videos that have minimal temporal resolution (needed for 3d convolutions), picking up both behavioral and structural elements of cell mixtures to correlate different proportions of cells with their behaviors. After implementing the algorithms on GPU using Theano/Lasagne, the results demonstrated that dCNNs outperform vanilla CNNs, particularly for distinguishing the proportions of cells that exist in given timelapse snippets.

Cell Pattern Recognition

Individual Project, Pattern Recognition, Biology

In this paper, I lay out a support vector machine (SVM) model with probabilistic classifications using Gabor wavelet and Haralick texture features of cells that can be used to determine cell structure and behavior in timelapse images. This model can be used to determine cell growth patterns and cell differentiation rates in real time for noninvasive clinical evaluations of cell culture.

DeepRock: Music Generation with AI

Team Project, NLU, Deep Learning, Music

We create a canonical encoding for multi-instrument MIDI songs into natural language, then use deep NLP techniques such as character LSTM variants to compose rock music that surpasses the prior state of the art and is competitive with certain pieces of music composed by human rock bands. We further define a neural network architecture for learning multi-instrument music generation in concert.

Mind Over Wheels

Team Project, Signal Processing, Machine Learning, Kinematic Control

We control the Barrett WAM robotic arm using standard control algorithms and signal processing routines of raw EEG signals collected from the brain (and other auxiliary signals such as accelerometer data). We hope to extend this approach to controlling wheelchairs in the future, and we hope to add functionality using machine learning to separate left-handed and right-handed movements using the EEG signals from different sides of the brain.

Hepatitis B Decision Tool

Team Project, Web Design

The purpose of this tool is to help policy makers evaluate/calculate the costs and benefits of chronic hepatitis CHB treatment strategies in different international settings. The potential usefulness of this information for health policy and planning is in assessing whether current intervention strategies represent an efficient use of scarce resources, and which of the potential additional interventions that are not yet or fully implemented, should be given priority on the grounds of cost-effectiveness. For example, the tool will help the policy maker assess at what negotiated antiviral drug cost would treatment be highly cost-effective.

Human Blastocyst Viewer

Individual Project, Data Viz

This is a 3D reconstruction of the early and late stage human blastocyst based on single cell gene expression profiles of 89 genes in 241 single cells. Location of each single cell is based on projection of the first 3 principal components onto a unit sphere.

Network Analysis for College Football Rankings

Team Project, Network Analysis, Algorithms, Optimization

Using win probabilities, centrality measures, Bradley-Terry rankings, and loop detection and removal, we construct an optionally weighted directed graph representing a season of college football and derive from it a ranking of the college football teams for that season. We use weekly prediction to evaluate our ten implemented ranking algorithms and compare performance with the AP Poll. We find that our modified BeatPaths algorithm outperforms all the other algorithms we tried, weighted algorithms outperform their unweighted counterparts, and our garbage time weighting scheme slightly outperforms a margin of victory weighting scheme. Notably, many of our network-based algorithms outperform the AP poll, which is traditionally considered to be the go-to metric for comparing college football teams.

Predicting Buzzfeed Reactions

Team Project, NLU, Deep Learning

In this paper, we explore how to predict the emotional response distribution of an article with text and images, based solely on the article’s text. In particular, we establish a new data corpus of Buzzfeed articles using data scraping tools. We run Bayesian estimators with dependency parses and neural networks on this dataset and find that while dependency parses with linear models (Jensen-Shannon (J-S) distance 0.319, lower is better) do not improve over our baseline prediction (J-S distance 0.286), the recurrent neural network (RNN) is able to significantly outperform the baseline prediction (J-S distance 0.248).