Data Science

Dr. Arun Padakandla
University of Tennessee

Abstract : Recognizing patterns from data crucially relies on the availability of true empirical counts from databases. However, the need to preserve privacy of individuals motivates perturbation of original databases, thus distorting empirical counts. We are thus led to the problem of designing efficient database perturbation schemes that distort empirical counts minimally while being robust to privacy attacks. Quantifying the robustness to privacy attacks via the framework of Differential Privacy and leveraging tools from discrete geometry, I characterize the minimal distortion as a function of the privacy guarantee.

In the second part, I present my recent findings on the problem of learning from quantum data. We consider a scenario wherein data is encoded onto quantum states and the corresponding labels are stored in classical registers. The problem of designing an identifying an optimal predictor or classifier manifests as a problem of identifying a suitable measurement from a concept class. We exploit ideas of joint measurability to propose a new ERM rule and characterize its sample complexity.

Lastly, I briefly speak about my current research and teaching activities and highlight some of my past research findings.