Eugenia Iofinova

Research Interests

I study bias in machine learning systems (and by extension, algorithmic fairness, interpretable ML, and generalization) in the context of on-device deep neural network models.

Scientific CV

Education

Since September 2020: PhD student in Computer Science/Data Science at the Institute of Science and Technology Austria, Klosterneuburg; Supervisor: Dan Alistarh; proposed thesis project: Understanding and mitigating bias in deployed models

September 2002-June 2006: California Institute of Technology, Pasadena, CA; B.S. Mathematics; GPA: 3.6/4.0 (Honors)

Research Experience

Institute of Science and Technology Austria Vienna, Austria
Research Internship, Alistarh Group June 2020-September 2020
Project: Weight and gradient pruning for image recognition models.

University of Vienna Vienna, Austria
Pandemic Forecasting Task Force May 2020-Aug 2020
Project: Forecasting the spread of the COVID-19 pandemic in Austria under different mitigation scenarios (available at epimath.at).

Center for Molecular Medicine (CeMM) Vienna, Austria
Research Internship, Menche Group April 2019-June 2020
Project 1: DataDiVR: Interactive 3D Virtual Reality viewer for large biological networks. (paper)
Project 2: RadiPOP: Outcome predictions for Portal Hypertension from CT scans. (patent submitted;
paper in preparation)

UCLA Institute for Pure and Applied Mathematics (IPAM) Los Angeles, CA
Co-supervisor, Summer undergraduate research internship June 2017 - Aug 2017
Project: Simulate human errors in document labeling and create optimal strategies for minimizing
prediction errors due to these given a limited rating budget. (Team size: 4 students)

California Institute of Technology Pasadena, CA
Summer undergraduate research internship, Aschbacher group July 2005-September 2005
Project: Rewrite some finite group theory results to fusion systems framework.

California Institute of Technology Pasadena, CA
Summer undergraduate research fellowship, Alvarez group July 2004-September 2004
Project: Comparison of absentee voting protocols in democratic nations.

Professional Experience

Google Los Angeles, CA
Software Engineer / Senior Software Engineer December 2014-February 2019
- Built first-of-their-kind deep learning-based binary and taxonomic classifiers for predicting subject
matter and sensitive content in text and video ads.
- Developed and launched ML fairness initiative in ads to correct biased misclassifications.

Castlight Health San Francisco, CA
Web Developer/ML Engineer/Strategic Analytics November 2009-December 2014
Built and deployed models for pricing services and predicting patient behavioral changes.

Upward Bound San Francisco, CA
Teacher (summer school in chemistry and physics) March 2009 - July 2009

Susquehanna International Group Philadelphia, PA
Algorithmic Trader May 2006 - Jan 2009

Publications

Published

Alexandra Peste, Eugenia Iofinova, Adrian Vladu, Dan Alistarh. AC/DC: Alternating Compressed/ DeCompressed Training of Deep Neural Networks. accepted at NeurIPS 2021; preprint available at
http://arxiv.org/abs/2106.12379.

Sebastian Pirch, Felix Müller, Eugenia Iofinova, Julia Pazmandi, Christiane VR Hütter, Martin Chiettini, Celine Sin, Kaan Boztug, Iana Podkosova, Hannes Kaufmann, Jörg Menche. The VRNetzer platform enables interactive network analysis in Virtual Reality. Nature communications, 23 April 2021.

Preprints

Eugenia Iofinova*, Alexandra Peste*, Mark Kurtz, Dan Alistarh. How Well Do Sparse ImageNet Features Transfer? submitted for review; preprint available at http://arxiv.org/abs/2111.13445.

Eugenia Iofinova*, Nikola Konstantinov*, Christoph H. Lampert. FLEA: Provably Fair Multisource Learning from Unreliable Training Data. pending submission; preprint available at
http://arxiv.org/abs/2106.11732.

Manu Eder, Joachim Hermisson, Michal Hledik, Christiane Hütter, Eugenia Iofinova, Rahul Pisupati, Jitka Polechova, Gemma Puixeu, Srdjan Sarikas, Benjamin Wölfl, Claudia Zimmermann. EpiMath Austria SEIR: A COVID-19 Compartment Model for Austria. preprint available at www.epimath.at/static/EpiMathAustria_SEIR_documentation.pdf