I am a PhD student in Statistics at Columbia University, working with Liam Paninski. I am interested in machine learning and information theory, the geometry of exponential families, graphical models, state-space models, functional data analysis / spatial statistics, structured non-iid data (e.g. networks), low-rank methods (Independent Component/Subspace Analysis, manifold learning), sparsity, and techniques to make computation tractable.
academic historyColumbia University 2010-, PhD in Statistics
University of British Columbia 2008-2010, MSc in Computer Science
Carnegie Mellon University 2006-2008, programmer for HCII, researcher at Machine Learning Department
Universiteit van Amsterdam 2003-2005, MSc in Logic at ILLC
Bucknell University 1997-2001, B.S. in Mathematics and Computer Science
why so many places, so many degrees?
conferences and summer schoolsIPAMGSS 2007 ICML/UAI 2008 SFI Summer School 2009
NIPS 2008, 2009 CogSci 2008, 2009.
technical tips (this is largely for myself)Collaborative Q&A sites: mathoverflow, math.stackexchange, CrossValidated
If interested in computationally-intensive problems, learn Julia, learn distributed programming.
Learn R. Use IRC. Visit the #R channel on FreeNode and get your R questions answered. Emacs users can use IRC by doing "M-x erc".
Use my code: R-helpers
statistical puzzlesSuppose you have (Xi,Yi) are i.i.d. from a bivariate Gaussian with correlation ρ, which we are interested in estimating. Suppose further that marginally, X and Y are standard normal. Unfortunately, the data manager was using a spreadsheet and accidentally sorted X without sorting Y, losing the information of which X goes with which Y. Is the data useless?
papers all publications Google ScholarIdentification of gene modules using a generative model for relational data (PDF, slides) - UBC Master's thesis (2010), supervised by Jennifer Bryan.
Discovering Cyclic Causal Models by ICA (UAI2008) (paper, video lecture with slides) extends LiNGAM to discover cyclic models; The non-Gaussian model leads to a finer level of identifiability than what can be achieved in the Gaussian case (e.g. by Richardson's CCD), and allows us to relax the faithfulness assumption. We prove theorems about identifiability, specifically about when a unique model can be identified.
(draft) Upper-Bounding Proof Length with the Busy Beaver (2008) (PDF) - This note presents a Chaitin-esque result. I derive an (uncomputable) upper bound on the length of the shortest proof of any given statement, as a function of the length of the statement; and briefly discuss implications. Mathematically trivial, but original (to the best of my knowledge). Could possibly be useful if we ever have good estimates of BB for n large enough to encode an interesting question.
see all papers
tutorials- Independent Component Analysis (ICA) (slides) Introduces ICA, and tackles some very common misconceptions, 30 minutes.
- Introduction to Kolmogorov Complexity (with Liliana Salvador) (slides), 45 minutes.
- Introduction to Machine Learning and Bayesian inference (slides), 45 minutes.
some things I likeargument mapping, bikes, bluegrass, contact improvisation, DreamWidth, emacs, functional programming, GiveWell, infoviz, musical instruments, open data, Quantified Self, wikis.
food for thought"You and Your Research", by Richard Hamming
"Why People Are Irrational about Politics", by Michael Huemer
"Why I defend scoundrels", by Yvain
Paul Graham: "How to do Philosophy", "Why nerds are unpopular"
LessWrong: Applause Lights
"Illusion of Transparency: Why No One Understands You"
Ribbonfarm: "A Big Little Idea Called Legibility"
Ben Goldacre: "The Information Architecture of Medicine is Broken"
blogsAndrew Gelman - Statistical Modeling, Causal Inference, and Social Science
Cosma Shalizi - Three-Toed Sloth
Cathy O' Neil - mathbabe
Peter Gray - Freedom to Learn