Subhodh Kotekal

Massachusetts Institute of Technology
2 previous affiliations
7Machine learning papers tracked by Lacuna
Is this you?

Research Overview

Subhodh Kotekal focuses on high-dimensional statistics, score-based generative models, and signal detection. His research establishes minimax rates and optimal estimators for complex statistical problems, particularly in settings where the underlying signal is sparse or the noise variance is unknown. Currently, he investigates the statistical optimality of diffusion models and the estimation of distributional functionals under smoothness constraints, bridging classical statistical theory with modern generative AI paradigms.

Research Trajectory

Kotekal's early research was situated in computational neuroscience, specifically examining the drivers of neuronal response variability in the visual cortex. In Recurrent interactions can explain the variance in single trial responses, he investigated whether local correlations between neighboring neurons or global variables better explain the activity of V1 neurons. By evaluating pairwise correlations of neuronal activity, this work contributed to the understanding of how recurrent interactions shape single-trial responses, providing a local alternative to global descriptions of visual encoding.

By 2021, his focus transitioned toward the fundamental challenges of high-dimensional statistics, particularly the "curse of dimensionality" where variables far outnumber observations. This period involved developing methods for Adaptive Testing for Sparse and Dense High-Dimensional Signals. His work addressed the dilemma of whether a signal is sparse, affecting only a few variables, or dense, affecting many variables weakly. This led to investigations into Adaptive Thresholding for Sparse Signal Recovery, where the objective is to identify signals without prior knowledge of sparsity levels or signal strength. During this phase, he also explored Exploiting Correlation for Signal Detection, demonstrating that redundancy in correlated sensors can be leveraged to extract faint signals from overwhelming noise.

In 2024, Kotekal's research deepened into compound decision theory and the computational-statistical gaps inherent in high-dimensional inference. He addressed the problem of Statistical Indistinguishability from Moment Matching, exploring the limits of what is mathematically possible versus what is computationally feasible. A significant contribution during this time was Variance estimation in compound decision theory under boundedness, where he established the sharp minimax rate for variance estimation in the normal means model when the means are bounded. He proposed a rate-optimal estimator that leverages a variational representation of noise variance in terms of the data distribution's cumulants, a technique related to his broader interest in Noise Suppression via Higher Order Cumulants.

Concurrently, he investigated Minimax Signal Detection in Sparse Additive Models, establishing nonasymptotic separation rates for signals where univariate component functions belong to a generic reproducing kernel Hilbert space. This research revealed a non-trivial interaction between sparsity and the choice of function space, further refining the theory of adaptive testing in high-dimensional settings.

Current Directions

Kotekal's recent work focuses on the theoretical underpinnings of score-based generative AI and diffusion models. He has contributed to understanding the statistical rates of score estimation in From optimal score matching to optimal sampling. This research demonstrates how the score matching risk—the estimation of the score function of a forward diffusion process—controls the total variation distance between the law of a generated sample and the ground truth distribution.

His research also extends to the Distributional Smoothness Constraints in Statistical Learning Theory. In Sharp Optimality of Simple, Plug-in Estimation of the Fisher Information of a Smoothed Density, he proved that simple plug-in estimators are minimax optimal for Fisher information in Gaussian-smoothed settings. This finding is significant because it allows for the estimation of mutual information and entropy at the ideal parametric rate, providing a solid theoretical foundation for information-theoretic tools used in deep learning.

Most recently, in Diffusion models are optimal for hypothesis testing, Kotekal explored the utility of the score-based paradigm for statistical hypothesis testing. He framed the task as detecting positive Fisher divergence between a noised null distribution and a noised, unknown data distribution. This work addresses the conceptual puzzle of why diffusion models, which simplify sampling by adding noise, remain effective for testing even as the null and alternative hypotheses become harder to distinguish at higher noise levels.

Representative Work

Recurrent interactions can explain the variance in single trial responses This work evaluates local and global descriptions of neuronal response variability in mouse V1. By computing pairwise correlations, the study demonstrates that local interactions between neighboring neurons significantly shape single-trial responses, offering a more granular explanation than global stimulus-driven models.

Variance estimation in compound decision theory under boundedness Addressing the difficulty of estimating noise variance in high-dimensional data with one observation per parameter, this paper establishes the sharp minimax rate of variance estimation under the assumption of bounded means. The proposed estimator achieves this rate by estimating a specific number of cumulants and utilizing a variational representation of the noise variance.

Minimax Signal Detection in Sparse Additive Models This paper establishes the minimax separation rate for detecting sparse additive signals in high-dimensional settings. The results are nonasymptotic and applicable to general reproducing kernel Hilbert spaces, highlighting how the choice of function space interacts with sparsity to determine detection limits.

From optimal score matching to optimal sampling As a second author, Kotekal contributed to establishing the statistical rates for score estimation in diffusion models. The work provides theoretical bounds on the total variation distance between generated and ground truth distributions, grounded in the risk associated with score matching.

Sharp Optimality of Simple, Plug-in Estimation of the Fisher Information of a Smoothed Density This research proves that simple plug-in estimators are minimax optimal for calculating the Fisher information of Gaussian-smoothed densities. This result simplifies the estimation of entropy and mutual information, which typically requires complex debiasing, and provides a theoretical basis for information-theoretic applications in deep learning.

Diffusion models are optimal for hypothesis testing This paper investigates the potential of the score-based paradigm of diffusion models for statistical hypothesis testing. It formulates the problem as detecting Fisher divergence and addresses the limitations and optimality of diffusion-based testing as noise levels increase.

Profile updated

See something wrong? Report a mistake