Contact Information- note that I am now at Carnegie Mellon

Jordan Rodu
Department of Statistics
Baker Hall
Carnegie Mellon University
Pittsburgh, PA 15123
jrodu@stat.cmu.edu

Papers

Locating recombination hot spots in genomic sequences through the singular value decomposition Submitted - with Shane Jensen


Spectral estimation of multiscale factorial hidden markov models Submitted- with Joao Sedoc, Dean Foster, Lyle Ungar


Jordan Rodu, Dean P. Foster, Weichen Wu, Lyle H. UngarUsing Regression for Spectral Estimation of HMMs. SLSP 2013: 212-223


Paramveer S. Dhillon, Jordan Rodu, Michael Collins, Dean P. Foster, Lyle H. Ungar: Spectral Dependency Parsing with Latent Variables. EMNLP-CoNLL 2012: 205-213


Paramveer S. Dhillon, Jordan Rodu, Dean P. Foster, Lyle H. Ungar: Using CCA to improve CCA: A new spectral method for estimating vector models of words. ICML 2012


Dean P. Foster, Jordan Rodu, Lyle H. Ungar: Spectral dimensionality reduction for HMMs. CoRR abs/1203.6130 (2012)


Papers in Progress

A Multidimensional Cross Correlogram for two multi-trial multivariate signals with Robert Kass

Abstract.This paper explores a method for generating a cross correlogram for multivariate time series in a repeated trial structure. Applications to simultaneously recorded neural data from two different brain regions.


Spectral Estimation of HMMs with a continuous output distribution with Dean Foster

Abstract. In this paper we consider spectral estimation of a discrete time, discrete hidden state space HMM with a continuous output distribution. While spectral estimation has been extended to the continuous output distribution HMM case, its estimation requires embedding the HMM into a reproducing kernel Hilbert space, a technology that requires storage of large amounts of training data and a significant amount of time to run per step. We provide a more natural extension of the classic spectral estimation technology to the continous output case that relieves the space and time requirements of the RKHS technology.


Spectral Estimation of a hierarchical HMM - with Dean Foster and Lyle Ungar

Abstract. In this paper we extend spectral estimation to hierarchical HMMs. We consider an HMM whose output is either discrete or continuous, and whose parameters change slowly over time according to a higher level HMM.


An MDP clustering of neurons by their hidden state paths

Abstract. We use a Dirichlet Process to cluster neurons according to their hidden state trajectories. Specifically, we model neuron microcuircutry under the assumption that a neuron's firing rate is governed by an HMM, and we make the assumption that neurons in the same cluster follow the same hidden state path. We assume that HMM parameters are shared accross clusters.


Spectral forward-backward probabilities and their use as features with Joao Sedoc, Dean Foster, and Lyle Ungar

Abstract.This paper proposes a spectral method for estimating spectral quantities analogous to forward-backward probabilities in standard HMM analysis. We explore these quantities as features in other algorithms, and as the foundation for a hierarchical spectral HMM.

Collaborations with Scientists

Characterizing the dysfunction of microcircuitry of neurons in mouse brains that exhibit obsessive com- pulsive disorder– with Suzanne Ahmari’s lab (Department of Psychiatry, University of Pittsburgh)


Detecting excess synchrony in mouse brains that exhibit symptoms of Parkinson’s disease– with Taylor Pospisil (student in Statistics), Valerie Venture, Max G’Sell, and Aryn Gittis’ lab (Biological Sciences, Carnegie Mellon University)