About Me


Greetings!


My name is Madalina (Ina) Fiterau, I am an Assistant Professor at the College of Information and Computer Sciences at UMass Amherst, having started there in the Fall of 2018. See my UMass profile page.

Previously, I was a Postdoctoral Scholar in the Computer Science Department at Stanford University and part of the Mobilize Center, working with Professors Chris Ré and Scott Delp.

I have obtained my PhD in Machine Learning from CMU in the Fall of 2015. My PhD thesis advisor is Dr. Artur Dubrawski, head of the Auton Lab. I obtained my master's degree in Machine Learning from CMU in 2012.


NOTE: This webpage will be kept updated with publications from my PhD and Postdoc for a while, but will ultimately be replaced with a faculty webpage, which is currently under construction.
In the meantime, please read about my past and ongoing projects below.

I aim to build intelligent systems that aid decision-making by extracting compact models and salient representations from multimodal, heterogeneous data.
Topics: dimensionality reduction, interpretability, transfer learning, hybrid models, vision, pattern detection, time series analysis.
I am especially interested in the use of machine learning systems for medical applications and have co-organized several NIPS workshops on this topic: MLCDA NIPS 2013, ML4CHG NIPS 2014, ML for Health 2016, ML for Health 2017.

I started off as a software engineer, having received my BEng from Politehnica Timisoara, Romania in 2009. Designing and coding software systems is something I enjoy, which is one of the reasons I've twice interned at Google, working with Jonathan Scott (summer of 2011) and Parisa Haghani (summer of 2013). I have also completed a research internship at Microsoft Research Cambridge, working with Peter Kontschieder and Samuel Rota-Buló.


Curriculum Vitae

Full resume. Compact resume.


Awards



Speaking Engagements*


*invited presentations, outside of contributed talks at conferences

  • Invited panelist at the Learning with Limited Data Workshop at NIPS, 9th of December 2017.
    Panel topic: Limited Labeled Data in Medical Imaging.

  • Oxford, September 26th 2017. Host: Mihaela van der Schaar.
    Topic: hybrid machine learning models for heterogeneous, multimodal data.

  • University of Southern California, May 26th 2017. Hosts: Dave Kale, Greg ver Steeg, Aram Galstyan
    Topic: learning representations from biomedical time series in the presence of structured information.

  • University of Washington, May 22nd 2017. Host: Noah Smith.
    Topic: learning representations from biomedical time series in the presence of structured information.


Active Projects (Stanford)


  • Learning cross-modal representations from biomedical signals of various types.
    [representation learning; deep learning; time series; images]

  • Application: Activity recognition from triaxial accelerometer data.
    [ensemble learning; time series]

  • Application: Predicting surgical outcomes for children with cerebral palsy from clinical indicators and gait analysis.
    [classification; decision support; gait kinematics]

  • Application: Predicting the incidence of running injuries based on walking and running patterns.
    [classification; pattern detection; gait kinematics]

  • Application: Pathology detection in heart MRIs from the UK Biobank.
    [deep learning; image processing; weak supervision]


Past Projects (CMU)


  • Visualization for Informative Projection Retrieval (VIPR). Explores the idea of building small ensembles of low-dimensional components which are applicable to significant subsets of data, such that any given sample can be handled using one of these sub-models or a sparse mixture of them. The methods I'm developing optimize the assignment of samples to sub-models while reducing the dimensionality of the solvers in the ensemble. We introduced a regression-based algorithm which identifies informative projections by optimizing over a matrix of point-wise loss estimators, learning compact models for classification, clustering, and regression. The framework also works in an active learning setting.
    [ensemble methods; structured sparsity; feature selection; projection retrieval; query-specific models]

  • Application: Artifact adjudication for health status alerts. This is a collaborative project with a team from the University of Pittsburgh: Dr. Gilless Clermont, Dr. Marilyn Hravnak, Dr. Michael Pinsky. We work with a cardio-respiratory monitoring system designed to process multiple vital signs indicative of the health status of ICU patients. The system issues an alert whenever some form of instability requires attention, but in practice, a substantial fraction of these alerts are artifacts. A subset of the data has been manually labeled and the aim is to use that subset to determine which of of the unlabeled samples are worth the experts' attention.
    [active learning; dimensionality reduction; ensemble methods]

  • Application: Detection and adjudication of radiation levels from vehicle scans. One of the many projects under way at the Auton Lab, this presents a multiclass discrimination problem where expert feedback is available. The models need to be interpretable. Another interesting aspect of this dataset is that the training and test samples come from different distributions.
    [multilabel classification; interpretable models; transductive learning]

  • Vital Sign Monitoring System that predicts clinical alarms and detects anomalous patterns in patient vital signs. This project uses the waveform and treatment data in the Mimic II dataset.
    [dynamic Bayesian Networks; variational methods]

  • Explanation-Oriented Partitioning - the task of finding sets of high-accuracy regions in low-dimensional subspaces. The framework we developed uses any given hypothesis class and enables visualization of the classification process.
    [ensemble methods; query-specific models]

  • Finding explanations in datasets by exploring areas in low-dimensional projections of the feature space. I've implemented some heuristic approaches as well as LPs which explicitly minimize entropy-based objective functions.
    [feature selection; information theory; linear programming]


Teaching Assistantships



Publications


Below are all my publications by year. View publication list by year and category.


2017



2016



2015



2014



2013



2012



2011



2010



2009


Contact

Madalina Fiterau-Brostean
PhD in Machine Learning
Postdoctoral Researcher
Stanford University
318 Campus Drive
James H. Clark Center, S331
Stanford, CA 94305
Skype phone: 412-336-8465
Twitter: Mfiterau


To obtain my e-mail address, run the following script in the python interpreter:
''.join([x for s in map(lambda(x): x[0:x.find('ad') if x.find('ad')>=0 else len(x)],list(['madalina','fiterau'])) + zip(list(['@','.','.']),list(['cs','cmu','edu'])) for x in s])

In case you don't have python handy (seriously?), my e-mail address is username@domain where my username is on the URL of the page (after ~) and the domain is cs.cmu.edu.
Design by Free Flash Templates | XHTML | CSS


Latest


9th of December 2017
Heart MRI preliminary work presented at the Medical Imaging NIPS Workshop.


8th of December 2017
Machine Learning for Health Workshop at NIPS 2017.


7th of December 2017
Poster on Surgical outcome prediction shown at the Interpretable ML Symposium.


18th of August 2017
ShortFuse presented at the Machine Learning for Healthcare Conference.


18th of August 2017
ShortFuse presented at the Machine Learning for Healthcare Conference.


18th of May 2017
Invited participant at the LATTICE Symposium.



Older


December 2016
Contextual LSTM presented at the WiML Workshop and the NIPS RNN Symposium.


30th of October 2016
Invited participant at the Rising Stars Workshop.


14th of July 2016
VIPR demo presented at IJCAI 2016.
View website | camera ready | poster.


12th of July 2016
Deep Neural Decision Forests presented at IJCAI 2016.
View camera ready | slides.


22nd of February 2016
Expert study for RIPR presented at SCCM 2016.
View abstract | slides.


14th of December 2015
Deep Neural Decision Forests wins Marr prize for Best Paper at ICCV 2015.
View camera ready | spotlight | poster.



27th of January 2015
The ActiveRIPR framework was presented at AAAI 2015.
view camera ready | poster.



18th of January 2015
The research on archetyping artifacts from alerts in vital sign monitoring data was presented at SCCM.
view slides.



12th of December 2014
The 2nd Workshop on Machine Learning for Clinical Data Analysis, Healthcare and Genomics took place at NIPS.
Many thanks to our speakers, panelists, poster presenters, and PC members!



27th of October 2014
Completed thesis proposal, Discovering Compact and Informative Structures through Data Partitioning.
Many thanks to my committee members and collaborators!
view slides.



10th of December 2013
The Machine Learning for Clinical Data Analysis Workshop took place at NIPS.
Many thanks to our speakers, panelists, poster presenters and PC members!



4th of December 2013
The RIPR algorithm presented at ICMLA 2013.
view camera ready and slides.



30th of September 2013
The paper Automatic Identification of Alarm Artifacts in Monitoring Critically Ill Patients presented at ESICM-LIVES.
view electronic poster.



20th of June 2013
RIPR models for artifact adjudication shown at the ICML-HEALTH Workshop.
view camera ready | poster.



3rd of December 2012
The RECIP algorithm was shown at NIPS 2012.
view camera ready | poster.



3rd of December 2012
'Feature-Task Co-clustering Regression' was presented at WiML 2012.
view abstract | poster.



21st of March 2012
Data Analysis Project. Committee: Prof. Artur Dubrawski, Prof. Geoff Gordon, Prof. Jeff Schneider
view slides pptx|pdf.



12th of December 2011
the algorithm `Explanation-Oriented Partitioning' was presented at the WiML workshop in Granada.
view slides.



7th of December 2011
the paper `Real-time Adaptive Monitoring of Vital Signs for Clinical Alarm Preemption' presented at the iSDS Annual Conference in Atlanta.
view poster.



29th of November 2011
Aiming for the Moon, an outline of the Lander visual registering project was presented during the ML Journal Club.