Krzysztof Chalupka: Research Interests

I am interested in understanding causation and connecting causal discovery and inference methods to large-scale machine learning frameworks. Most of my work at Caltech has been on Causal Feature Learning -- a theoretical and algorithmic framework for automatic, unbiased formation of causally well-defined variables from large data.

During my PhD I've also worked on understanding vision-driven behavior of the fruit fly, helped build a robot inspired by this behavior and performed single-cell in-vivo recordings from the macaque visual cortex.

Publications

Unsupervised Discovery of El Nino Using Causal Feature Learning on Microlevel Climate Data
Krzysztof Chalupka, Tobias Bischoff, Pietro Perona and Frederick Eberhardt
UAI 2016, plenary presentation.





Multi-Level Cause-Effect Systems
Krzysztof Chalupka, Pietro Perona and Frederick Eberhardt
AISTATS 2016.





Visual Causal Feature Learning
Krzysztof Chalupka, Pietro Perona and Frederick Eberhardt
UAI 2015, accepted for plenary presentation. Code and Tetrad Counterexamples





Generalized Regressive Motion: a Visual Cue to Collision
Krzysztof Chalupka, Michael Dickinson, Pietro Perona (2014)
Bioinspiration and Biomimetics, Code and Supplementary Movies





A Framework for Evaluating Approximation Methods for Gaussian Process Regression
Krzysztof Chalupka, Christopher K. I. Williams and Iain Murray
JMLR 2013. Code and Data.




Previous Research

Causality in Vision.
With Pietro Perona and Frederick Eberhardt.
Formal causal reasoning methods have been clarified and popularized in late 1990s by Judea Pearl and Peter Spirtes. I'm working on creating causally meaningful visual models, as well as extracting causal (as opposed to purely correlational) information from images and videos.


Low-energy Multiclass Classification.
With Ron Appel and Qualcomm.
Starting September 2014, Qualcomm's Innovation Fellowship supports our work on learning features efficiently and cheaply under various cost constraints. Whether it's energy consumption, time cost or bandwidth constraints that you care about, we can learn you a classifier that you can afford and that does as well as possible!

Other Past Projects

Robot Navigation with Generalized Regressive Motion.
Thanks to Caltech's Summer Undergraduate Research Fellowship (SURF) Program, I was able to mentor a high school student Amrit Rau during Summer 2014. Amrit built a robot that used an Android phone's camera and computing power for navigation. The robot computes the optical flow and detects impending collisions using an algorithm developed by Pietro Perona, Michael Dickinson, and me. Amrit's technical report

Active Learning and Brain-Computer Interfaces.
With Andreas Krause and Alex Roper.
As a Caltech class project, Alex and I created an algorithm that, based on Wordnet semantic trees, adaptively chooses questions to ask, trying to guess the word a user is thinking of. Our algorithm was based on Bayesian Optimization. At some point we realized that there is a an EEG-readable brainwave frequency that relates to the semantic distance between what a person thinks and perceives (Google N400 if you're curious, that's the frequency). This allowed our class project to develop into full-fledged research when we purchased an EEG machine using Caltech's Housner Grant.

Bayesian Protein Structure Determination.
With Michael Habeck.
X-ray crystallography, nuclear magnetic resonance spectroscopy, and electron microscopy can provide a wealth of information about proteins that can be used to determine their structure. I worked on a Bayesian framework for protein structure determination . We used state-of-the-art replica-exchange MCMC and John Skilling's nested sampling algorithm.

Krzysztof Chalupka
web statistics