arXiv subjects

Abhirup Ghosh

Publications and source records attributed to Abhirup Ghosh.

At least 19 recordsLinked to original sources

arXiv2026

NEO: No-Optimization Test-Time Adaptation through Latent Re-Centering

Test-Time Adaptation (TTA) methods are often computationally expensive, require a large amount of data for effective adaptation, or are brittle to hyperparameters. Based on a theoretical foundation of the geometry of the latent space, we are able to significantly improve the alignment between source and distribution-shifted samples by re-centering target data embeddings at the origin. This insight motivates NEO -- a hyperparameter-free fully TTA method, that adds no significant compute compared to vanilla inference. NEO is able to improve the classification accuracy of ViT-Base on ImageNet-C from 55.6% to 59.2% after adapting on just one batch of 64 samples. When adapting on 512 samples NEO beats all 7 TTA methods we compare against on ImageNet-C, ImageNet-R and ImageNet-S and beats 6/7 on CIFAR-10-C, while using the least amount of compute. NEO performs well on model calibration metrics and additionally is able to adapt from 1 class to improve accuracy on 999 other classes in ImageNet-C. On Raspberry Pi and Jetson Orin Nano devices, NEO reduces inference time by 63% and memory usage by 9% compared to baselines. Our results based on 3 ViT architectures and 4 datasets show that NEO can be used efficiently and effectively for TTA.

cs.LG↗

arXiv2026

Disentangling Reasoning in Large Audio-Language Models for Ambiguous Emotion Prediction

Speech emotion recognition plays an important role in various applications. However, most existing approaches predict a single emotion label, oversimplifying the inherently ambiguous nature of human emotional expression. Recent large audio-language models show promise in generating richer outputs, but their reasoning ability for ambiguous emotional understanding remains limited. In this work, we reformulate ambiguous emotion recognition as a distributional reasoning problem and present the first systematic study of ambiguity-aware reasoning in LALMs. Our framework comprises two complementary components: an ambiguity-aware objective that aligns predictions with human perceptual distributions, and a structured ambiguity-aware chain-of-thought supervision that guides reasoning over emotional cues. Experiments on IEMOCAP and CREMA-D demonstrate consistent improvements across SFT, DPO, and GRPO training strategies.

cs.SD↗

arXiv2026

E-BATS: Efficient Backpropagation-Free Test-Time Adaptation for Speech Foundation Models

Speech Foundation Models encounter significant performance degradation when deployed in real-world scenarios involving acoustic domain shifts, such as background noise and speaker accents. Test-time adaptation (TTA) has recently emerged as a viable strategy to address such domain shifts at inference time without requiring access to source data or labels. However, existing TTA approaches, particularly those relying on backpropagation, are memory-intensive, limiting their applicability in speech tasks and resource-constrained settings. Although backpropagation-free methods offer improved efficiency, existing ones exhibit poor accuracy. This is because they are predominantly developed for vision tasks, which fundamentally differ from speech task formulations, noise characteristics, and model architecture, posing unique transferability challenges. In this paper, we introduce E-BATS, the first Efficient BAckpropagation-free TTA framework designed explicitly for speech foundation models. E-BATS achieves a balance between adaptation effectiveness and memory efficiency through three key components: (i) lightweight prompt adaptation for a forward-pass-based feature alignment, (ii) a multi-scale loss to capture both global (utterance-level) and local distribution shifts (token-level) and (iii) a test-time exponential moving average mechanism for stable adaptation across utterances. Experiments conducted on four noisy speech datasets spanning sixteen acoustic conditions demonstrate consistent improvements, with 4.1%-13.5% accuracy gains over backpropagation-free baselines and 2.0-6.4 times GPU memory savings compared to backpropagation-based methods. By enabling scalable and robust adaptation under acoustic variability, this work paves the way for developing more efficient adaptation approaches for practical speech processing systems in real-world environments.

cs.LG↗

arXiv2026

A meta inspiral-merger-ringdown consistency test of general relativity with gravitational wave signals from compact binaries

The observation of gravitational waves from compact binary coalescences is a promising tool to test the validity of general relativity (GR) in a highly dynamical strong-field regime. There are now a variety of tests of GR performed on the observed compact binary signals. In this paper, we propose a new test of GR that compares the results of these individual tests. This meta inspiral-merger-ringdown consistency test (IMRCT) involves inferring the final mass and spin of the remnant black hole obtained from the analyses of two different tests of GR and checking for consistency. If there is a deviation from GR, we expect that different tests of GR will recover different values for the final mass and spin, in general. We check the performance of the meta IMRCT using a standard set of null tests used in various gravitational-wave analyses: the original IMRCT, parameterized phasing tests (TIGER and FTI) and the modified dispersion test. However, the meta IMRCT is applicable to any tests of GR that infer the initial masses and spins or the final mass and spin, including ones that are applied to binary neutron star or neutron star--black hole signals. We apply the meta IMRCT to simulated quasi-circular GR and non-GR binary black hole (BBH) signals as well as to eccentric BBH signals in GR (analyzed with quasicircular waveforms). We find that the meta IMRCT gives consistency with GR for the quasi-circular GR signals and picks up a deviation from GR in the other cases, as do other tests. In some cases, the meta IMRCT finds a significant GR deviation for a given pair of tests (and specific testing parameters) while the individual tests do not, showing that it is more sensitive than the individual tests to certain types of deviations. In addition, we also apply this test to a few selected real compact binary signals and find them consistent with GR.

gr-qc↗

arXiv2025

BoTTA: Benchmarking on-device Test Time Adaptation

The performance of deep learning models depends heavily on test samples at runtime, and shifts from the training data distribution can significantly reduce accuracy. Test-time adaptation (TTA) addresses this by adapting models during inference without requiring labeled test data or access to the original training set. While research has explored TTA from various perspectives like algorithmic complexity, data and class distribution shifts, model architectures, and offline versus continuous learning, constraints specific to mobile and edge devices remain underexplored. We propose BoTTA, a benchmark designed to evaluate TTA methods under practical constraints on mobile and edge devices. Our evaluation targets four key challenges caused by limited resources and usage conditions: (i) limited test samples, (ii) limited exposure to categories, (iii) diverse distribution shifts, and (iv) overlapping shifts within a sample. We assess state-of-the-art TTA methods under these scenarios using benchmark datasets and report system-level metrics on a real testbed. Furthermore, unlike prior work, we align with on-device requirements by advocating periodic adaptation instead of continuous inference-time adaptation. Experiments reveal key insights: many recent TTA algorithms struggle with small datasets, fail to generalize to unseen categories, and depend on the diversity and complexity of distribution shifts. BoTTA also reports device-specific resource use. For example, while SHOT improves accuracy by $2.25\times$ with $512$ adaptation samples, it uses $1.08\times$ peak memory on Raspberry Pi versus the base model. BoTTA offers actionable guidance for TTA in real-world, resource-constrained deployments.

cs.LG↗

arXiv2025

Benchmarking Federated Machine Unlearning methods for Tabular Data

Machine unlearning, which enables a model to forget specific data upon request, is increasingly relevant in the era of privacy-centric machine learning, particularly within federated learning (FL) environments. This paper presents a pioneering study on benchmarking machine unlearning methods within a federated setting for tabular data, addressing the unique challenges posed by cross-silo FL where data privacy and communication efficiency are paramount. We explore unlearning at the feature and instance levels, employing both machine learning, random forest and logistic regression models. Our methodology benchmarks various unlearning algorithms, including fine-tuning and gradient-based approaches, across multiple datasets, with metrics focused on fidelity, certifiability, and computational efficiency. Experiments demonstrate that while fidelity remains high across methods, tree-based models excel in certifiability, ensuring exact unlearning, whereas gradient-based methods show improved computational efficiency. This study provides critical insights into the design and selection of unlearning algorithms tailored to the FL environment, offering a foundation for further research in privacy-preserving machine learning.

cs.LG↗

arXiv2024

FLea: Addressing Data Scarcity and Label Skew in Federated Learning via Privacy-preserving Feature Augmentation

Federated Learning (FL) enables model development by leveraging data distributed across numerous edge devices without transferring local data to a central server. However, existing FL methods still face challenges when dealing with scarce and label-skewed data across devices, resulting in local model overfitting and drift, consequently hindering the performance of the global model. In response to these challenges, we propose a pioneering framework called \textit{FLea}, incorporating the following key components: \textit{i)} A global feature buffer that stores activation-target pairs shared from multiple clients to support local training. This design mitigates local model drift caused by the absence of certain classes; \textit{ii)} A feature augmentation approach based on local and global activation mix-ups for local training. This strategy enlarges the training samples, thereby reducing the risk of local overfitting; \textit{iii)} An obfuscation method to minimize the correlation between intermediate activations and the source data, enhancing the privacy of shared features. To verify the superiority of \textit{FLea}, we conduct extensive experiments using a wide range of data modalities, simulating different levels of local data scarcity and label skew. The results demonstrate that \textit{FLea} consistently outperforms state-of-the-art FL counterparts (among 13 of the experimented 18 settings, the improvement is over $5\%$) while concurrently mitigating the privacy vulnerabilities associated with shared features. Code is available at https://github.com/XTxiatong/FLea.git

cs.LG↗

arXiv2023

Tests of general relativity in the nonlinear regime: a parametrized plunge-merger-ringdown gravitational waveform model

The plunge-merger stage of the binary-black hole coalescence, when the bodies' velocities reach a large fraction of the speed of light and the gravitational-wave luminosity peaks, provides a unique opportunity to probe gravity in the dynamical and nonlinear regime. How much do the predictions of general relativity differ from the ones in other theories of gravity for this stage of the binary evolution? To address this question, we develop a parametrized waveform model, within the effective-one-body formalism, that allows for deviations from general relativity in the plunge-merger-ringdown stage. As first step, we focus on nonprecessing-spin, quasicircular black hole binaries. In comparison to previous works, for each gravitational wave mode, our model can modify, with respect to general-relativistic predictions, the instant at which the amplitude peaks, the instantaneous frequency at this time instant, and the value of the peak amplitude. We use this waveform model to explore several questions considering both synthetic-data injections and two gravitational wave signals. In particular, we find that deviations from the peak gravitational wave amplitude and instantaneous frequency can be constrained to about $20\%$ with GW150914. Alarmingly, we find that GW200129_065458 shows a strong violation of general relativity. We interpret this result as a false violation, either due to waveform systematics (mismodeling of spin precession) or due to data-quality issues depending on one's interpretation of this event. This illustrates the use of parametrized waveform models as tools to investigate systematic errors in plain general relativity. The results with GW200129_065458 also vividly demonstrate the importance of waveform systematics and of glitch mitigation procedures when interpreting tests of general relativity with current gravitational wave observations.

gr-qc↗

arXiv2023

Cross-device Federated Learning for Mobile Health Diagnostics: A First Study on COVID-19 Detection

Federated learning (FL) aided health diagnostic models can incorporate data from a large number of personal edge devices (e.g., mobile phones) while keeping the data local to the originating devices, largely ensuring privacy. However, such a cross-device FL approach for health diagnostics still imposes many challenges due to both local data imbalance (as extreme as local data consists of a single disease class) and global data imbalance (the disease prevalence is generally low in a population). Since the federated server has no access to data distribution information, it is not trivial to solve the imbalance issue towards an unbiased model. In this paper, we propose FedLoss, a novel cross-device FL framework for health diagnostics. Here the federated server averages the models trained on edge devices according to the predictive loss on the local data, rather than using only the number of samples as weights. As the predictive loss better quantifies the data distribution at a device, FedLoss alleviates the impact of data imbalance. Through a real-world dataset on respiratory sound and symptom-based COVID-$19$ detection task, we validate the superiority of FedLoss. It achieves competitive COVID-$19$ detection performance compared to a centralised model with an AUC-ROC of $79\%$. It also outperforms the state-of-the-art FL baselines in sensitivity and convergence speed. Our work not only demonstrates the promise of federated COVID-$19$ detection but also paves the way to a plethora of mobile health model development in a privacy-preserving fashion.

cs.LG↗

arXiv2023

Black-hole ringdown as a probe of higher-curvature gravity theories

Detecting gravitational waves from coalescing compact binaries allows us to explore the dynamical, nonlinear regime of general relativity and constrain modifications to it. Some of the gravitational-wave events observed by the LIGO-Virgo Collaboration have sufficiently high signal-to-noise ratio in the merger, allowing us to probe the relaxation of the remnant black hole to its final, stationary state - the so-called black-hole ringdown, which is characterized by a set of quasinormal modes. Can we use the ringdown to constrain deviations from general relativity, as predicted by several of its contenders? Here, we address this question by using an inspiral-merger-ringdown waveform model in the effective-one-body formalism, augmented with a parametrization of the ringdown based on an expansion in the final black hole's spin. We give a prescription on how to include in this waveform model, the quasinormal mode frequencies calculated on a theory-by-theory basis. In particular, we focus on theories that modify general relativity by higher-order curvature corrections, namely, Einstein-dilaton-Gauss-Bonnet (EdGB), dynamical Chern-Simons (dCS) theories, and cubic- and quartic-order effective-field-theories (EFT) of general relativity. We use this parametrized waveform model to measure the ringdown properties of the two loudest ringdown signals observed so far, GW150914 and GW200129. We find that while EdGB theory cannot be constrained with these events, we can place upper bounds on the fundamental lengthscale of cubic- ($\ell_{\rm cEFT} \leqslant 38.2$ km) and quartic-order ($\ell_{\rm qEFT} \leqslant 51.3$ km) EFTs of general relativity, and of dCS gravity ($\ell_{\rm dCS} \leqslant 38.7$ km). The latter result is a concrete example of a theory presently unconstrained by inspiral-only analyses which, however, can be constrained by merger-ringdown studies with current gravitational-wave data.

gr-qc↗

arXiv2022

Constraining extra dimensions using observations of black hole quasi-normal modes

The presence of extra dimensions generically modify the spacetime geometry of a rotating black hole, by adding an additional hair, besides the mass $M$ and the angular momentum $J$, known as the `tidal charge' parameter, $β$. In a braneworld scenario with one extra spatial dimension, the extra dimension is expected to manifest itself through -- (a) negative values of $β$, and (b) modified gravitational perturbations. This in turn would affect the quasi-normal modes of rotating black holes. We numerically solve the perturbed gravitational field equations using the continued fractions method and determine the quasi-normal mode spectra for the braneworld black hole. We find that increasingly negative values of $β$ correspond to a diminishing imaginary part of the quasi-normal mode, or equivalently, an increasing damping time. Using the publicly available data of the properties of the remnant black hole in the gravitational wave signal GW150914, we check for consistency between the predicted values (for a given $β$) of the frequency and damping time of the least-damped $\ell=2,m=2$ quasi-normal mode and measurements of these quantities using other independent techniques. We find that it is highly unlikely for the tidal charge, $β\lesssim -0.05$, providing a conservative limit on the tidal charge parameter. Implications and future directions are discussed.

gr-qc↗

arXiv2022

Summary of Tests of General Relativity with GWTC-3

Observations of gravitational waves (GWs) by the advanced LIGO--Virgo detectors provide us with ground breaking opportunities to test predictions of Einstein's theory of general relativity (GR) in the strong field regime. In this article, we summarise the nine tests of GR performed on the new GW signals included in the third GW transient catalog, GWTC-3. These tests include overall and self-consistency checks of the signal with the data; tests of the GW generation, propagation and polarizations; and probes of the nature of the remnant object by testing the BH ringdown hypothesis and searching for post-merger echoes. The results from the new events are combined with those previously published wherever possible. We do not find any statistically significant deviation from GR and set the most stringent bounds yet on possible departures from theory.

gr-qc↗

arXiv2022

Tests of General Relativity with Gravitational-Wave Observations using a Flexible--Theory-Independent Method

We perform tests of General Relativity (GR) with gravitational waves (GWs) from the inspiral stage of compact binaries using a theory-independent framework, which adds generic phase corrections to each multipole of a GR waveform model in frequency domain. This method has been demonstrated on LIGO-Virgo observations to provide stringent constraints on post-Newtonian predictions of the inspiral and to assess systematic biases that may arise in such parameterized tests. Here, we detail the anatomy of our framework for aligned-spin waveform models. We explore the effects of higher modes in the underlying signal on tests of GR through analyses of two unequal-mass, simulated binary signals similar to GW190412 and GW190814. We show that the inclusion of higher modes improves both the precision and the accuracy of the measurement of the deviation parameters. Our testing framework also allows us to vary the underlying baseline GR waveform model and the frequency at which the non-GR inspiral corrections are tapered off. We find that to optimize the GR test of high-mass binaries, comprehensive studies would need to be done to determine the best choice of the tapering frequency as a function of the binary's properties. We also carry out an analysis on the binary neutron-star event GW170817 to set bounds on the coupling constant $α_0$ of Jordan-Fierz-Brans-Dicke gravity. We take two plausible approaches; in the first \emph{theory-agnostic} approach we find a bound $α_0 \lesssim 2\times 10^{-1}$ from measuring the dipole-radiation for different neutron-star equations of state, while in the second \emph{theory-specific} approach we obtain $α_0 \lesssim 4\times 10^{-1}$, both at $68\%$ credible level. These differences arise mainly due to different statistical hypotheses used for the analysis.

gr-qc↗

arXiv2022

Continual and Sliding Window Release for Private Empirical Risk Minimization

It is difficult to continually update private machine learning models with new data while maintaining privacy. Data incur increasing privacy loss -- as measured by differential privacy -- when they are used in repeated computations. In this paper, we describe regularized empirical risk minimization algorithms that continually release models for a recent window of data. One version of the algorithm uses the entire data history to improve the model for the recent window. The second version uses a sliding window of constant size to improve the model, ensuring more relevant models in case of evolving data. The algorithms operate in the framework of stochastic gradient descent. We prove that even with releasing a model at each time-step over an infinite time horizon, the privacy cost of any data point is bounded by a constant $ε$ differential privacy, and the accuracy of the output models are close to optimal. Experiments on MNIST and Arxiv publications data show results consistent with the theory.

cs.LG↗

arXiv2022

Investigating the relation between gravitational wave tests of general relativity

Gravitational wave observations of compact binary coalescences provide precision probes of strong-field gravity. There is thus now a standard set of null tests of general relativity (GR) applied to LIGO-Virgo detections and many more such tests proposed. However, the relation between all these tests is not yet well understood. We start to investigate this by applying a set of standard tests to simulated observations of binary black holes in GR and with phenomenological deviations from GR. We consider four types of tests: residuals, inspiral-merger-ringdown consistency, parameterized phasing (two varieties), and modified dispersion relation. We also check the consistency of the unmodeled reconstruction of the waveforms with the waveform recovered using GR templates. These tests are applied to simulated observations similar to GW150914 with both large and small deviations from GR and similar to GW170608 just with small deviations from GR. We find that while very large deviations from GR are picked up with high significance by almost all tests, more moderate deviations are picked up by only a few tests, and some deviations are not recognized as GR violations by any test at the moderate signal-to-noise ratios we consider. Moreover, the tests that identify various deviations with high significance are not necessarily the expected ones. In particular, the parameterized tests recover PN test parameters much closer to zero than their true values in some cases. Additionally, we find that of the GR deviations we consider, the residuals test is only able to detect extreme deviations from GR that no longer look like binary black hole coalescences in GR. The reconstruction comparison shows more promise for detecting relatively small GR deviations in an unmodeled framework, at least for high-mass systems.

gr-qc↗

arXiv2021

Population inference of spin-induced quadrupole moments as a probe for non-black hole compact binaries

Gravitational-wave (GW) measurements of physical effects such as spin-induced quadrupole moments can distinguish binaries consisting of black holes from non-black hole binaries. While these effects may be poorly constrained for single-event inferences with the second-generation detectors, combining information from multiple detections can help uncover features of non-black hole binaries. The spin-induced quadrupole moment has specific predictions for different types of compact objects, and a generalized formalism must consider a population where different types of compact objects co-exist. In this study, we introduce a hierarchical mixture-likelihood formalism to estimate the {\it fraction of non-binary black holes in the population}. We demonstrate the applicability of this method using simulated GW signals injected into Gaussian noise following the design sensitivities of the Advanced LIGO Advanced Virgo detectors. We compare the performance of this method with a traditionally-followed hierarchical inference approach. Both the methods are equally effective to hint at inhomogeneous populations, however, we find the mixture-likelihood approach to be more natural for mixture populations comprising compact objects of diverse classes. We also discuss the possible systematics in the mixture-likelihood approach, caused by several reasons, including the limited sensitivity of the second-generation detectors, specific features of the astrophysical population distributions, and the limitations posed by the waveform models employed. Finally, we apply this method to the LIGO-Virgo detections published in the second GW transient catalog (GWTC-2) and find them consistent with a binary black hole population within the statistical precision.

gr-qc↗

arXiv2021

Constraining cosmological extra dimensions with gravitational wave standard sirens: from theory to current and future multi-messenger observations

The propagation of gravitational waves (GWs) at cosmological distances offers a new way to test the gravitational interaction at the largest scales. Many modified theories of gravity, usually introduced to explain the observed acceleration of the universe, can be probed in an alternative and complementary manner with respect to standard electromagnetic (EM) observations. In this paper we consider a homogeneous and isotropic cosmology with extra spatial dimensions at large scales, which represents a simple phenomenological prototype for extra-dimensional modified gravity cosmological models. By assuming that gravity propagates through the higher-dimensional spacetime, while photons are constrained to the usual four dimensions of general relativity, we derive from first principles the relation between the luminosity distance measured by GW detectors and the one inferred by EM observations. We then use this relation to constrain the number of cosmological extra dimensions with the binary neutron star event GW170817 and the binary black hole merger GW190521. We further provide forecasts for the Laser Interferometer Space Antenna (LISA) by simulating multi-messenger observations of massive black hole binary (MBHB) mergers. This paper extends and updates previous analyses which crucially neglected an additional redshift dependency in the GW-EM luminosity distance relation which affects results obtained from multi-messenger GW events at high redshift, in particular constraints expected from LISA MBHBs.

gr-qc↗

arXiv2021

Constraints on quasi-normal-mode frequencies with LIGO-Virgo binary-black-hole observations

The no-hair conjecture in General Relativity (GR) states that a Kerr black hole (BH) is completely described by its mass and spin. As a consequence, the complex quasi-normal-mode (QNM) frequencies of a binary-black-hole (BBH) ringdown can be uniquely determined by the mass and spin of the remnant object. Conversely, measurement of the QNM frequencies could be an independent test of the no-hair conjecture. This paper extends to spinning BHs earlier work that proposed to test the no-hair conjecture by measuring the complex QNM frequencies of a BBH ringdown using parameterized inspiral-merger-ringdown waveforms in the effective-one-body formalism, thereby taking full advantage of the entire signal power and removing dependency on the predicted or estimated start time of the ringdown. Our method was used to analyze the properties of the merger remnants for BBHs observed by LIGO-Virgo in the first half of their third observing (O3a) run. After testing our method with GR and non-GR synthetic-signal injections in Gaussian noise, we analyze, for the first time, two BBHs from the first (O1) and second (O2) LIGO-Virgo observing runs, and two additional BBHs from the O3a run. We then provide joint constraints with published results from the O3a run. In the most agnostic and conservative scenario where we combine the information from different events using a hierarchical approach, we obtain, at $90\%$ credibility, that the fractional deviations in the frequency (damping time) of the dominant QNM are $δf_{220}=0.03^{+0.10}_{-0.09}$ ($δτ_{220}=0.10^{+0.44}_{-0.39}$), respectively, an improvement of a factor of $\sim 4$ ($\sim 2$) over the results obtained with our model in the LIGO-Virgo publication. The single-event most-stringent constraint to date continues to be GW150914 for which we obtain $δf_{220}=0.05^{+0.11}_{-0.07}$ and $δτ_{220}=0.07^{+0.26}_{-0.23}$.

gr-qc↗