arXiv subjects

Manjari Narayan

Publications and source records attributed to Manjari Narayan.

4 recordsLinked to original sources

arXiv2025

Causal Judge Evaluation: Calibrated Surrogate Metrics for LLM Systems

Measuring long-run LLM outcomes (user satisfaction, expert judgment, downstream KPIs) is expensive. Teams default to cheap LLM judges, but uncalibrated proxies can invert rankings entirely. Causal Judge Evaluation (CJE) makes it affordable to aim at the right target: calibrate cheap scores against a small oracle slice, then evaluate at scale with valid uncertainty. We treat surrogate validity as auditable: for each policy or deployment context, a small oracle audit tests whether the learned calibration remains mean-unbiased, turning an uncheckable identification condition into a falsifiable diagnostic. On 4,961 Chatbot Arena prompts comparing five policies with a 16x oracle/judge cost ratio, at a 5% oracle fraction CJE achieves 99% pairwise ranking accuracy at 14x lower cost; across all configurations (5-50% oracle, varying n), accuracy averages 94%. An adversarial policy fails the transport audit and is correctly flagged; in such cases CJE refuses level claims rather than reporting biased estimates. Key findings: naive confidence intervals on raw judge scores achieve 0% coverage (CJE: ~95%); importance-weighted estimators fail despite >90% effective sample size; and the Coverage-Limited Efficiency (CLE) bound and its TTC diagnostic explain why.

stat.ME↗

arXiv2015

Two Sample Inference for Populations of Graphical Models with Applications to Functional Connectivity

Gaussian Graphical Models (GGM) are popularly used in neuroimaging studies based on fMRI, EEG or MEG to estimate functional connectivity, or relationships between remote brain regions. In multi-subject studies, scientists seek to identify the functional brain connections that are different between two groups of subjects, i.e. connections present in a diseased group but absent in controls or vice versa. This amounts to conducting two-sample large scale inference over network edges post graphical model selection, a novel problem we call Population Post Selection Inference. Current approaches to this problem include estimating a network for each subject, and then assuming the subject networks are fixed, conducting two-sample inference for each edge. These approaches, however, fail to account for the variability associated with estimating each subject's graph, thus resulting in high numbers of false positives and low statistical power. By using resampling and random penalization to estimate the post selection variability together with proper random effects test statistics, we develop a new procedure we call $R^{3}$ that solves these problems. Through simulation studies we show that $R^{3}$ offers major improvements over current approaches in terms of error control and statistical power. We apply our method to identify functional connections present or absent in autistic subjects using the ABIDE multi-subject fMRI study.

stat.ME↗

arXiv2012

Anisotropic Nonlocal Means Denoising

It has recently been proved that the popular nonlocal means (NLM) denoising algorithm does not optimally denoise images with sharp edges. Its weakness lies in the isotropic nature of the neighborhoods it uses to set its smoothing weights. In response, in this paper we introduce several theoretical and practical anisotropic nonlocal means (ANLM) algorithms and prove that they are near minimax optimal for edge-dominated images from the Horizon class. On real-world test images, an ANLM algorithm that adapts to the underlying image gradients outperforms NLM by a significant margin.

math.ST↗

arXiv2011

Suboptimality of Nonlocal Means for Images with Sharp Edges

We conduct an asymptotic risk analysis of the nonlocal means image denoising algorithm for the Horizon class of images that are piecewise constant with a sharp edge discontinuity. We prove that the mean square risk of an optimally tuned nonlocal means algorithm decays according to $n^{-1}\log^{1/2+ε} n$, for an $n$-pixel image with $ε>0$. This decay rate is an improvement over some of the predecessors of this algorithm, including the linear convolution filter, median filter, and the SUSAN filter, each of which provides a rate of only $n^{-2/3}$. It is also within a logarithmic factor from optimally tuned wavelet thresholding. However, it is still substantially lower than the the optimal minimax rate of $n^{-4/3}$.

math.ST↗