Embedded Files
Understanding the geometry of flow-based generative models (In submission)
Understanding the geometry of flow-based generative models (In submission)
This project introduces a new straightness parameter for the flow-matching/ rectified flow image generative models that characterizes the performance of generated image quality. In particular, we obtain a novel bound on the Wasserstein convergence rate that is tested through extensive simulations on both synthetic and real data.
Understanding flight pattern through Fourier analysis (JSM award 2025 - Data expo, professional division)
Understanding flight pattern through Fourier analysis (JSM award 2025 - Data expo, professional division)
We propose a new statistical method to identify the flight pattern in the major US airports based on the Fourier spectral density of the flight arrival-departure pattern. In particular, we use the entropy of the spectral distribution to identify the level of frequency of the flight arrival and departure schedule to predict whether a specific airport is a hub for a particular airline or not.
Privatizing linear bandits using hard thresholding (AISTATS 2025):
Privatizing linear bandits using hard thresholding (AISTATS 2025):
Introduces FLIPHAT, a new private algorithm for the sparse linear bandit problem that achieves optimal performance under privacy in the high-dimensional sparse setting. This algorithm uses a privatized version of hard thresholding and a novel forgetting step within each episode to achieve an optimal privacy guarantee.
Private feature selection using MCMC (Neurips 2024):
Private feature selection using MCMC (Neurips 2024):
DP-BSS algorithm is a new MCMC-based private feature selection algorithm for the sparse linear regression problem. DP-BSS provably achieves computational efficiency over other private feature selection algorithms with more than 90% F1 score.
Thompson sampling in RL (ICML 2023):
Thompson sampling in RL (ICML 2023):
This project introduces a new Variational Bayes Thompson sampling algorithm (VBTS) for sparse and high-dimensional linear contextual bandits to understand and exploit the hidden sparsity structure of the context. VBTS achieves the SOTA performance in the Gravier data set for the online classification task with an accuracy of 82%.
Robust inference using new divergence (J. Math. Stat 2019):
Robust inference using new divergence (J. Math. Stat 2019):
A new divergence-based robust algorithm for parameter inference in the presence of outliers. This methodology uses the Bregman-divergence-based approach for optimizing a weighted likelihood-type loss function to mitigate the effect of outliers in the learning algorithm. The algorithm achieves the best MSE performance compared to the other existing algorithms.
Page updated
Report abuse