Ishani Karmarkar

I am a PhD student at Stanford University in the Institute for Computational and Mathematical Engineering. My research primarily focuses on the design and analysis of algorithms for problems arising in graph theory, machine learning, and optimization. I am advised by Aaron Sidford and Ellen Vitercik.

Previously, I earned a BS in Applied and Computational Mathematics with a Minor in Computer Science from the California Institute of Technology (Caltech), where I did undergraduate research with Bamdad Hosseini, Andrew Stuart, and Ashish Mahabal. I also had the opportunity to work as an intern at Facebook (now Meta), where I worked on a variety of machine learning projects.

Email  /  GitHub  /  Google Scholar  /  LinkedIn

Research Projects

From Large to Small Datasets: Size Generalization for Clustering Algorithm Selection

Clustering is a fundamental problem in machine learning. There are a variety of clustering algorithms in the literature; however, there is scant guidance on choosing a good clustering algorithm for a dataset at hand. In this paper we approach the problem of clustering algorithm selection using a semi-supervised framework. We present algorithms that use a small subset of the data to provably estimate the performance of three classic clustering algorithms: k-means++, k-centers, and single linkage hierarchical clustering on the full dataset. We support our theoretical results with experiments on several real-world clustering instances. (arxiv)

Towards Optimal Effective Resistance Estimation

Computing the effective resistance of nodes in a network (or, undirected graph) is a fundamental task with many applications, for instance, in graph machine learning, graph data analysis, and the design of efficient graph algorithms. In this paper we introduce new efficient algorithms for estimating effective resistances in well-connected graphs. We also prove improved time lower bounds for effective resistance estimation. In addition, we provide natural extensions of our work to related problems on symmetric diagonally dominant (SDD) and positive semi-definite (PSD) matrices. (arxiv)

Real-bogus classification for the Zwicky Transient Facility using deep learning

In this project, we present a convolutional neural network classifier for classifying real versus bogus data in astronomical images from the Zwicky Transients Facility (ZTF), a new robotic time-domain astronomical survey. (arxiv)

Design and source code from Jon Barron's website