arXiv subjects

Hari Shankar

Publications and source records attributed to Hari Shankar.

7 recordsLinked to original sources

arXiv2026

Mind the Gap: Pitfalls of LLM Alignment with Asian Public Opinion

Large Language Models (LLMs) are increasingly being deployed in multilingual, multicultural settings, yet their reliance on predominantly English-centric training data risks misalignment with the diverse cultural values of different societies. In this paper, we present a comprehensive, multilingual audit of the cultural alignment of contemporary LLMs including GPT-4o-Mini, Gemini-2.5-Flash, Llama 3.2, Mistral and Gemma 3 across India, East Asia and Southeast Asia. Our study specifically focuses on the sensitive domain of religion as the prism for broader alignment. To facilitate this, we conduct a multi-faceted analysis of every LLM's internal representations, using log-probs/logits, to compare the model's opinion distributions against ground-truth public attitudes. We find that while the popular models generally align with public opinion on broad social issues, they consistently fail to accurately represent religious viewpoints, especially those of minority groups, often amplifying negative stereotypes. Lightweight interventions, such as demographic priming and native language prompting, partially mitigate but do not eliminate these cultural gaps. We further show that downstream evaluations on bias benchmarks (such as CrowS-Pairs, IndiBias, ThaiCLI, KoBBQ) reveal persistent harms and under-representation in sensitive contexts. Our findings underscore the urgent need for systematic, regionally grounded audits to ensure equitable global deployment of LLMs.

cs.CL↗

arXiv2025

MetaGMT: Improving Actionable Interpretability of Graph Multilinear Networks via Meta-Learning Filtration

The growing adoption of Graph Neural Networks (GNNs) in high-stakes domains like healthcare and finance demands reliable explanations of their decision-making processes. While inherently interpretable GNN architectures like Graph Multi-linear Networks (GMT) have emerged, they remain vulnerable to generating explanations based on spurious correlations, potentially undermining trust in critical applications. We present MetaGMT, a meta-learning framework that enhances explanation fidelity through a novel bi-level optimization approach. We demonstrate that MetaGMT significantly improves both explanation quality (AUC-ROC, Precision@K) and robustness to spurious patterns, across BA-2Motifs, MUTAG, and SP-Motif benchmarks. Our approach maintains competitive classification accuracy while producing more faithful explanations (with an increase up to 8% of Explanation ROC on SP-Motif 0.5) compared to baseline methods. These advancements in interpretability could enable safer deployment of GNNs in sensitive domains by (1) facilitating model debugging through more reliable explanations, (2) supporting targeted retraining when biases are identified, and (3) enabling meaningful human oversight. By addressing the critical challenge of explanation reliability, our work contributes to building more trustworthy and actionable GNN systems for real-world applications.

cs.LG↗

arXiv2025

Sometimes the Model doth Preach: Quantifying Religious Bias in Open LLMs through Demographic Analysis in Asian Nations

Large Language Models (LLMs) are capable of generating opinions and propagating bias unknowingly, originating from unrepresentative and non-diverse data collection. Prior research has analysed these opinions with respect to the West, particularly the United States. However, insights thus produced may not be generalized in non-Western populations. With the widespread usage of LLM systems by users across several different walks of life, the cultural sensitivity of each generated output is of crucial interest. Our work proposes a novel method that quantitatively analyzes the opinions generated by LLMs, improving on previous work with regards to extracting the social demographics of the models. Our method measures the distance from an LLM's response to survey respondents, through Hamming Distance, to infer the demographic characteristics reflected in the model's outputs. We evaluate modern, open LLMs such as Llama and Mistral on surveys conducted in various global south countries, with a focus on India and other Asian nations, specifically assessing the model's performance on surveys related to religious tolerance and identity. Our analysis reveals that most open LLMs match a single homogeneous profile, varying across different countries/territories, which in turn raises questions about the risks of LLMs promoting a hegemonic worldview, and undermining perspectives of different minorities. Our framework may also be useful for future research investigating the complex intersection between training data, model architecture, and the resulting biases reflected in LLM outputs, particularly concerning sensitive topics like religious tolerance and identity.

cs.CY↗

arXiv2022

Leveraging Clinically Relevant Biometric Constraints To Supervise A Deep Learning Model For The Accurate Caliper Placement To Obtain Sonographic Measurements Of The Fetal Brain

Multiple studies have demonstrated that obtaining standardized fetal brain biometry from mid-trimester ultrasonography (USG) examination is key for the reliable assessment of fetal neurodevelopment and the screening of central nervous system (CNS) anomalies. Obtaining these measurements is highly subjective, expertise-driven, and requires years of training experience, limiting quality prenatal care for all pregnant mothers. In this study, we propose a deep learning (DL) approach to compute 3 key fetal brain biometry from the 2D USG images of the transcerebellar plane (TC) through the accurate and automated caliper placement (2 per biometry) by modeling it as a landmark detection problem. We leveraged clinically relevant biometric constraints (relationship between caliper points) and domain-relevant data augmentation to improve the accuracy of a U-Net DL model (trained/tested on: 596 images, 473 subjects/143 images, 143 subjects). We performed multiple experiments demonstrating the effect of the DL backbone, data augmentation, generalizability and benchmarked against a recent state-of-the-art approach through extensive clinical validation (DL vs. 7 experienced clinicians). For all cases, the mean errors in the placement of the individual caliper points and the computed biometry were comparable to error rates among clinicians. The clinical translation of the proposed framework can assist novice users from low-resource settings in the reliable and standardized assessment of fetal brain sonograms.

eess.IV↗

arXiv2022

Towards A Device-Independent Deep Learning Approach for the Automated Segmentation of Sonographic Fetal Brain Structures: A Multi-Center and Multi-Device Validation

Quality assessment of prenatal ultrasonography is essential for the screening of fetal central nervous system (CNS) anomalies. The interpretation of fetal brain structures is highly subjective, expertise-driven, and requires years of training experience, limiting quality prenatal care for all pregnant mothers. With recent advancement in Artificial Intelligence (AI), specifically deep learning (DL), assistance in precise anatomy identification through semantic segmentation essential for the reliable assessment of growth and neurodevelopment, and detection of structural abnormalities have been proposed. However, existing works only identify certain structures (e.g., cavum septum pellucidum, lateral ventricles, cerebellum) from either of the axial views (transventricular, transcerebellar), limiting the scope for a thorough anatomical assessment as per practice guidelines necessary for the screening of CNS anomalies. Further, existing works do not analyze the generalizability of these DL algorithms across images from multiple ultrasound devices and centers, thus, limiting their real-world clinical impact. In this study, we propose a DL based segmentation framework for the automated segmentation of 10 key fetal brain structures from 2 axial planes from fetal brain USG images (2D). We developed a custom U-Net variant that uses inceptionv4 block as a feature extractor and leverages custom domain-specific data augmentation. Quantitatively, the mean (10 structures; test sets 1/2/3/4) Dice-coefficients were: 0.827, 0.802, 0.731, 0.783. Irrespective of the USG device/center, the DL segmentations were qualitatively comparable to their manual segmentations. The proposed DL system offered a promising and generalizable performance (multi-centers, multi-device) and also presents evidence in support of device-induced variation in image quality (a challenge to generalizibility) by using UMAP analysis.

eess.IV↗

arXiv2014

Enhanced Precision Through Multiple Reads for LDPC Decoding in Flash Memories

Multiple reads of the same Flash memory cell with distinct word-line voltages provide enhanced precision for LDPC decoding. In this paper, the word-line voltages are optimized by maximizing the mutual information (MI) of the quantized channel. The enhanced precision from a few additional reads allows FER performance to approach that of full-precision soft information and enables an LDPC code to significantly outperform a BCH code. A constant-ratio constraint provides a significant simplification in the optimization with no noticeable loss in performance. For a well-designed LDPC code, the quantization that maximizes the mutual information also minimizes the frame error rate in our simulations. However, for an example LDPC code with a high error floor caused by small absorbing sets, the MMI quantization does not provide the lowest frame error rate. The best quantization in this case introduces more erasures than would be optimal for the channel MI in order to mitigate the absorbing sets of the poorly designed code. The paper also identifies a trade-off in LDPC code design when decoding is performed with multiple precision levels; the best code at one level of precision will typically not be the best code at a different level of precision.

cs.IT↗

arXiv2012

LDPC Decoding with Limited-Precision Soft Information in Flash Memories

This paper investigates the application of low-density parity-check (LDPC) codes to Flash memories. Multiple cell reads with distinct word-line voltages provide limited-precision soft information for the LDPC decoder. The values of the word-line voltages (also called reference voltages) are optimized by maximizing the mutual information (MI) between the input and output of the multiple-read channel. Constraining the maximum mutual-information (MMI) quantization to enforce a constant-ratio constraint provides a significant simplification with no noticeable loss in performance. Our simulation results suggest that for a well-designed LDPC code, the quantization that maximizes the mutual information will also minimize the frame error rate. However, care must be taken to design the code to perform well in the quantized channel. An LDPC code designed for a full-precision Gaussian channel may perform poorly in the quantized setting. Our LDPC code designs provide an example where quantization increases the importance of absorbing sets thus changing how the LDPC code should be optimized. Simulation results show that small increases in precision enable the LDPC code to significantly outperform a BCH code with comparable rate and block length (but without the benefit of the soft information) over a range of frame error rates.

cs.IT↗