headshot.jpg

Michał Dereziński

Email: mderezin@ucsc.edu
Website: http://users.soe.ucsc.edu/~mderezin

I am currently in the fifth year of my Ph.D. at the University of California, Santa Cruz, advised by professor Manfred Warmuth. In my thesis research, I develop efficient data sampling techniques with applications to learning theory and optimization. Prior to UCSC, I completed Master's degrees in mathematics and computer science at the University of Warsaw. I also interned at a variety of Silicon Valley research labs (e.g. Microsoft, Yahoo, eBay), working on projects ranging from online learning to large-scale distributed optimization.


Education and Research

Ph.D. student, Computer Science, University of California, Santa Cruz.

Subsampling for Ridge Regression via Regularized Volume Sampling
Proposed a new efficient procedure for solving a ridge regression task over a large unlabeled dataset, when computing labels is expensive. This method, called regularized volume sampling, is fast, easy to implement, and offers strong statistical guarantees.
2017.  arXiv:1710.05110

Batch-Expansion Training: An Efficient Optimization Paradigm
Proposed a parameter-free batch optimization method running on a gradually expanding dataset, which outperforms standard batch and stochastic approaches in large-scale settings.
2017.  arXiv:1704.06731

Discovering Surprising Documents with Context-Aware Word Representations
Developed a information-theoretic model for discovering surprise in text, that outperforms deep learning in a practical task.
IUI 2018 (to appear).  PDF

Unbiased estimates for linear regression via volume sampling
Showed that volume sampling provides an unbiased estimate of the least squares solution from a subset of labels, which is close to the optimum solution up to a multiplicative factor. Also, proposed an efficient algorithm for volume sampling.
NIPS 2017.  arXiv:1705.06908  Poster  Spotlight

Anticipating Concept Drift in Online Learning
Developed and analyzed online optimization algorithms that dynamically adapt to concept drift.
Workshop paper at NIPS 2015.  PDF

The limits of squared Euclidean distance regularization
Proved a lower bound for the loss of learning algorithms that regularize with squared Euclidean distance, for a certain family of learning problems. Proposed and experimentally tested a conjecture extending the result to deep learning.
NIPS 2014.  PDF  Spotlight


M.S. in Computer Science, University of Warsaw.

On Generating Concept Hierarchies with Fuzzy Data
Developed and tested a new algorithm for generating concept hierarchies from fuzzy data, that is more efficient and accurate than existing approaches.
Master's Thesis, 2012.  PDF

Active Semi-Supervised Concept Hierarchy Refinement
Developed and tested a new algorithm for refining concept hierarchies generated from unsupervised data, by actively querying a human expert.
Workshop paper at LAWS 2012.  PDF


M.S. in Mathematics, University of Warsaw.

Isomorphic Properties of Function Space BV on Simply Connected Planar Sets
Master's Thesis, 2013.  PDF

Wzrost w Języku Struktur Zgrubnych [Growth in Terms of Coarse Structures]
Bachelor's Thesis, 2011.  PDF


Experience

Research Intern, Microsoft Research, Sunnyvale, CA
Developed a new optimization framework combining the benefits of stochastic and batch methods.

Intern Research Scientist, Yahoo Inc., Sunnyvale, CA
Developed optimization algorithms for recommendation systems that dynamically adapt to changing environment.

Intern Research Scientist, eBay Inc., San Jose, CA
Developed and implemented an unsupervised recommendation system based on topic modeling, for finding interesting products.

Software engineer Intern, Nvidia, Santa Clara, CA.

Intern, Interdisciplinary Centre for Mathematical and Computational Modelling, Warsaw, Poland.
Researched algorithms related to concurrent programming.


Other Interests

Classical Music.
Playing and composing piano music.
Favorite composer: Johann Sebastian Bach.

Sports.
Mountain biking, hiking, tennis, table tennis.