## Marco Cusumano-Towner |

I am a PhD student in Electrical Engineering and Computer Science at MIT, working in the MIT Probabilistic Computing Project on abstractions for stochastic computation and techniques for measuring the runtime accuracy of approximate inference algorithms. I am supported by an NDSEG fellowship. From 2012 to 2015 I was an early-stage employee at a Sequoia-backed startup operating in clinical cancer diagnostics, where I developed computational infrastructure and applied machine learning algorithms for DNA-sequencing-based genetic testing. Prior to this, I completed a Master's in Computer Science at Stanford University, where I was funded by the NSF Graduate Research Fellowship Program. At Stanford, I worked with Serafim Batzoglou and Anshul Kundaje on machine learning algorithms for predictive modeling of gene expression. I did my undergraduate studies in Electrical Engineering and Computer Science at UC Berkeley, where I worked with Pieter Abbeel on probabilistic robotics and computer vision for robotics.

Cusumano-Towner M, Radul A, Wingate D, Mansinghka V. arXiv preprint arXiv:1704.04977 (2017)

Cusumano-Towner M, Mansinghka V. arXiv preprint arXiv:1612.04759 (2016).

Cusumano-Towner M, Mansinghka V. Presented at the NIPS 2016 Advances in Approximate Bayesian Inference Workshop.

Cusumano-Towner M, Mansinghka V. 2016. arXiv preprint arXiv:1606.00068 (2016).

Cusumano-Towner M, Saad F. Video presentation of joint work of the MIT Probabilistic Computing Project at the 2015 Future Programming Workshop.

Kyriazopoulou-Panagiotopoulou S,

Kyriazopoulou-Panagiotopoulou S, Cusumano-Towner M, Batzoglou S, Kundaje A. 2013 Genome Informatics CSHL Meeting.

Kyriazopoulou-Panagiotopoulou S, Cusumano-Towner M, Batzoglou S, Kundaje A. 2013 NIPS Workshop on Machine Learning in Computational Biology (MLCB).

We used EMR data from Stanford Hospital to construct a probabilistic infection network which we used for simulation experiments.

We formulate a cloth simulator suitable for robotics applications as a convex optimization problem, as well as a HMM-based framework for tracking the state of cloth objects through a sequence of manipulations. We apply these methods to autonomously and reliably perform the challenging task of bringing an unidentified crumpled article of clothing into a desired configuration.

Maitin-Shepard J,

We present a novel vision-based cloth grasp point detection algorithm and demonstrate it on the end-to-end task of autonomously folding towels. The system successfully folds 50 out of 50 previously-unseen towels.

I had just taken CS228T at Stanford and was excited to try out some of the methods. In this project I used blocked Gibbs sampling to do inference in a dynamic infection network with SIR (susceptible-infectious-resistant) dynamics.

For use in the research on predictive modeling of gene expression, I made the following writeups and derivations to help myself understand the relationship between boosting and optimization, and derived some efficient algorithms for our particular problem.

An algorithm for efficiently finding the K-Nearest-Neighbors of a test example, when the training and test data lie on some discrete lattice in feature space. This was used to efficiently run KNN on the genes x conditions lattice of examples used in the predictive modeling of gene expression research.

Final project for scientific computing course at Stanford.

For use in the research on predictive modeling of gene expression, I made the following writeups and derivations to help myself understand the relationship between boosting and optimization, and derived some efficient algorithms for our particular problem.

An algorithm for efficiently finding the K-Nearest-Neighbors of a test example, when the training and test data lie on some discrete lattice in feature space. This was used to efficiently run KNN on the genes x conditions lattice of examples used in the predictive modeling of gene expression research.

Stanford CS 448 (Data Visualization) final project in which I visualized gene sets using interactive dimensionality reduction for a visual version of enrichment analysis.

Stanford CS 273a (Computational Tour of Human Genome) final project in which we learned to predict epigenetic marker signals across the genome from a small set of these markers. We investigated how well we could predict epigenetic states using a subset of the markers to impute the missing signals.

Stanford CS 229 (Machine Learning) final project in which I applied several clustering algorithms to time-series gene expression data.

Berkeley CS281a (Fall 2009, Statistical Learning Theory) final project in which we trained a 1-dimensional CRBM to perform unsupervised feature learing of helicopter trajectory features, for use in segmentation of helicopter trajectories into different maneuvers.

In this brief project I was experimenting with iteratively learning a distance metric that agreed with a user's implicit internal metric, based on continuous user feedback in a dimensionality reduction setting. The algorithm takes gradient descent steps in a (diagonal) metric learning objective simultaneously with an interactive MDS.

A program written in C++ that superimposes a human figure onto a nondeterministic version of Conway's Game of Life, and modifies the probabilities based on location in the image.