Cyrus Rashtchian (UCSD) - Foundations of Data Science: Algorithms, Models, Explainability

Foundations of Data Science: Algorithms, Models, Explainability

Watch via live stream

Building a theory for data science involves formulating new theoretical frameworks for important applications, as well as developing efficient and reliable solutions for associated computational challenges. Central themes of my research include new models and algorithms for bioinformatics and trustworthy machine learning. In this talk, I first describe my work on trustworthy machine learning, where I will present a new model for explainable k-means clustering based on small decision trees and a new algorithm for finding a tree-based clustering with provably low cost. This work is the first to identify an unsupervised learning problem where explainable-by-design algorithms do not suffer from a large loss in their effectiveness. Turning to DNA data storage, I will provide an overview of this exciting, emerging technology. It promises orders of magnitude improved density and longevity compared to existing storage media. However, efficiently retrieving data that has been stored in DNA requires solving many interesting theoretical and practical problems. I will survey my contributions in this area, including efficient DNA synthesis methods, a distributed clustering algorithm for edit distance, and new statistical reconstruction algorithms. Next, I will discuss how to reconstruct node-labeled trees when given samples from an appropriately defined deletion channel. This involves new combinatorial and statistical algorithms, and it showcases a difficult model where worst-case reconstruction is possible with a polynomial number of samples. Finally, I will share my plans for future work in the areas of statistical reconstruction, trustworthy machine learning, and applied algorithms more generally.

Host: Sanjay Krishnan