Hrishikesh Viswanath

LiDAR mapping as energy optimization over the curvature of a neural signed distance field

A survey of advances and challenges in data-driven Scientific ML

A mesh-free training scheme that amortizes the cost of Monte Carlo walks by using the Walk-on-Spheres algorithm to provide cheap, unbiased stochastic supervision for neural operators.

A sampling-based framework that evolves Lagrangian systems directly in physical space using data-driven neural PDE operators and learnable kernel ROM parameterization

We introduce a physics-informed regularization for offline GCRL derived from the Hamilton-Jacobi-Bellman (HJB) equation’s viscosity solution. By applying the Feynman-Kac theorem, we recast value estimation as a stable Monte Carlo expectation, ensuring geometric consistency and superior performance in high-dimensional navigation and manipulation tasks.

A PINO technique for training operators using Monte-Carlo Walk-on-Spheres

We introduce a resolution-independent neural operator architecture for synthetic frame generation that leverages global convolutions in the Fourier spectral domain via Fast Fourier Transform (FFT).

A decentralized graph neural operator approach for robust and scalable task allocation in heterogeneous UGV-UAV teams.

An Optical Flow-Integrated Markov Neural Operator that models pedestrian trajectory prediction as a Markovian process, eliminating the need for historical state storage.

We present a comprehensive survey and a modular toolkit of mathematical frameworks for quantifying and mitigating prediction bias in Large Language Models (LLMs).