arXiv subjects

Anirban Roy

Publications and source records attributed to Anirban Roy.

At least 55 records · Page 3Linked to original sources

arXiv2022

Detecting out-of-context objects using contextual cues

This paper presents an approach to detect out-of-context (OOC) objects in an image. Given an image with a set of objects, our goal is to determine if an object is inconsistent with the scene context and detect the OOC object with a bounding box. In this work, we consider commonly explored contextual relations such as co-occurrence relations, the relative size of an object with respect to other objects, and the position of the object in the scene. We posit that contextual cues are useful to determine object labels for in-context objects and inconsistent context cues are detrimental to determining object labels for out-of-context objects. To realize this hypothesis, we propose a graph contextual reasoning network (GCRN) to detect OOC objects. GCRN consists of two separate graphs to predict object labels based on the contextual cues in the image: 1) a representation graph to learn object features based on the neighboring objects and 2) a context graph to explicitly capture contextual cues from the neighboring objects. GCRN explicitly captures the contextual cues to improve the detection of in-context objects and identify objects that violate contextual relations. In order to evaluate our approach, we create a large-scale dataset by adding OOC object instances to the COCO images. We also evaluate on recent OCD benchmark. Our results show that GCRN outperforms competitive baselines in detecting OOC objects and correctly detecting in-context objects.

cs.CV↗

arXiv2022

Probing the circumgalactic medium with CMB polarization statistical anisotropy

As cosmic microwave background (CMB) photons traverse the Universe, anisotropies can be induced via Thomson scattering (proportional to the integrated electron density; optical depth) and inverse Compton scattering (proportional to the integrated electron pressure; thermal Sunyaev-Zel'dovich effect). Measurements of anisotropy in optical depth $τ$ and Compton $y$ parameter are imprinted by the galaxies and galaxy clusters and are thus sensitive to the thermodynamic properties of circumgalactic medium and intergalactic medium. We use an analytic halo model to predict the power spectrum of the optical depth ($ττ$), the cross-correlation between the optical depth and the Compton $y$ parameter ($τy$), as well as the cross-correlation between the optical depth and galaxy clustering ($τg$), and compare this model to cosmological simulations. We constrain the optical depths of halos at $z\lesssim 3$ using a technique originally devised to constrain patchy reionization at a much higher redshift range. The forecasted signal-to-noise ratio is 2.6, 8.5, and 13, respectively, for a CMB-S4-like experiment and a VRO-like optical survey. We show that a joint analysis of these probes can constrain the amplitude of the density profiles of halos to 6.5% and the pressure profile to 13%, marginalizing over the outer slope of the pressure profile. These constraints translate to astrophysical parameters related to the physics of galaxy evolution, such as the gas mass fraction, $f_{\rm g}$, which can be constrained to 5.3% uncertainty at $z\sim 0$, assuming an underlying model for the shape of the density profile. The cross-correlations presented here are complementary to other CMB and galaxy cross-correlations since they do not require spectroscopic galaxy redshifts and are another example of how such correlations are a powerful probe of the astrophysics of galaxy evolution.

astro-ph.CO↗

arXiv2022

iDECODe: In-distribution Equivariance for Conformal Out-of-distribution Detection

Machine learning methods such as deep neural networks (DNNs), despite their success across different domains, are known to often generate incorrect predictions with high confidence on inputs outside their training distribution. The deployment of DNNs in safety-critical domains requires detection of out-of-distribution (OOD) data so that DNNs can abstain from making predictions on those. A number of methods have been recently developed for OOD detection, but there is still room for improvement. We propose the new method iDECODe, leveraging in-distribution equivariance for conformal OOD detection. It relies on a novel base non-conformity measure and a new aggregation method, used in the inductive conformal anomaly detection framework, thereby guaranteeing a bounded false detection rate. We demonstrate the efficacy of iDECODe by experiments on image and audio datasets, obtaining state-of-the-art results. We also show that iDECODe can detect adversarial examples.

cs.LG↗

arXiv2021

MISA: Online Defense of Trojaned Models using Misattributions

Recent studies have shown that neural networks are vulnerable to Trojan attacks, where a network is trained to respond to specially crafted trigger patterns in the inputs in specific and potentially malicious ways. This paper proposes MISA, a new online approach to detect Trojan triggers for neural networks at inference time. Our approach is based on a novel notion called misattributions, which captures the anomalous manifestation of a Trojan activation in the feature space. Given an input image and the corresponding output prediction, our algorithm first computes the model's attribution on different features. It then statistically analyzes these attributions to ascertain the presence of a Trojan trigger. Across a set of benchmarks, we show that our method can effectively detect Trojan triggers for a wide variety of trigger patterns, including several recent ones for which there are no known defenses. Our method achieves 96% AUC for detecting images that include a Trojan trigger without any assumptions on the trigger pattern.

cs.CR↗

arXiv2021

Constraining reionization with the first measurement of the cross-correlation between the CMB optical-depth fluctuations and the Compton y-map

We propose a new reionization probe that uses cosmic microwave background (CMB) observations; the cross-correlation between fluctuations in the CMB optical depth which probes the integrated electron density, $δτ$, and the Compton $y$-map which probes the integrated electron pressure. This cross-correlation is much less contaminated than the $y$-map power spectrum by late-time cluster contributions. In addition, this cross-correlation can constrain the temperature of ionized bubbles while the optical-depth fluctuations and kinetic SZ effect can not. We measure this new observable using a Planck $y$-map as well as a map of optical-depth fluctuations that we reconstruct from Planck CMB temperature data. We use our measurements to derive a first CMB-only upper limit on the temperature inside ionized bubbles, $T_{\rm b}\lesssim 7.0\times10^5\,$K ($2\,σ$). We also present future forecasts, assuming a fiducial model with characteristic reionization bubble size $R_{\rm b}=5\,$Mpc and $T_{\rm b}=5\times10^4\,$K. The signal-to-noise ratio of the fiducial cross-correlation using a signal dominated PICO-like $y$-map becomes $\simeq7$ with CMB-S4 $δτ$ and $\simeq13$ with CMB-HD $δτ$. For the fiducial model, we predict that the CMB-HD $-$ PICO cross-correlation should achieve an accurate measurement of the reionization parameters; $T_{\rm b}\simeq 49800^{+4500}_{-5100}\,$K and $R_{\rm b}\simeq 5.09^{+0.66}_{-0.79}\,$Mpc. Since the power spectrum of the electron density fluctuations is constrained by the $δτ$ auto spectrum, the temperature constraints should be only weakly model-dependent on the details of the electron distributions and should be statistically representative of the temperature in ionized bubbles during reionization. This cross-correlation could, therefore, become an important observable for future CMB experiments.

astro-ph.CO↗

arXiv2021

Detecting OODs as datapoints with High Uncertainty

Deep neural networks (DNNs) are known to produce incorrect predictions with very high confidence on out-of-distribution inputs (OODs). This limitation is one of the key challenges in the adoption of DNNs in high-assurance systems such as autonomous driving, air traffic management, and medical diagnosis. This challenge has received significant attention recently, and several techniques have been developed to detect inputs where the model's prediction cannot be trusted. These techniques detect OODs as datapoints with either high epistemic uncertainty or high aleatoric uncertainty. We demonstrate the difference in the detection ability of these techniques and propose an ensemble approach for detection of OODs as datapoints with high uncertainty (epistemic or aleatoric). We perform experiments on vision datasets with multiple DNN architectures, achieving state-of-the-art results in most cases.

cs.LG↗

arXiv2021

Are all outliers alike? On Understanding the Diversity of Outliers for Detecting OODs

Deep neural networks (DNNs) are known to produce incorrect predictions with very high confidence on out-of-distribution (OOD) inputs. This limitation is one of the key challenges in the adoption of deep learning models in high-assurance systems such as autonomous driving, air traffic management, and medical diagnosis. This challenge has received significant attention recently, and several techniques have been developed to detect inputs where the model's prediction cannot be trusted. These techniques use different statistical, geometric, or topological signatures. This paper presents a taxonomy of OOD outlier inputs based on their source and nature of uncertainty. We demonstrate how different existing detection approaches fail to detect certain types of outliers. We utilize these insights to develop a novel integrated detection approach that uses multiple attributes corresponding to different types of outliers. Our results include experiments on CIFAR10, SVHN and MNIST as in-distribution data and Imagenet, LSUN, SVHN (for CIFAR10), CIFAR10 (for SVHN), KMNIST, and F-MNIST as OOD data across different DNN architectures such as ResNet34, WideResNet, DenseNet, and LeNet5.

cs.LG↗

arXiv2021

Revised estimates of CMB $B$-mode polarization induced by patchy reionization

The search for primordial gravitational waves through the $B$-mode polarization pattern in the CMB is one of the major goals of current and future CMB experiments. Besides foregrounds, a potential hurdle in this search is the anisotropic secondary $B$-mode polarization generated by the scattering of CMB photons off free electrons produced during patchy cosmological reionization. Robust predictions of these secondary anisotropies are challenging because of uncertainties in the reionization history. In this paper, we revise estimates of the reionization-induced $B$-mode signal by incorporating recent advances in the understanding of reionization through observations of the Lyman-$α$ forest. To derive these $B$-mode estimates, we use high-dynamic-range radiative transfer simulations of reionization that are calibrated to the Ly$α$ data. These simulations are also consistent with a variety of other high-redshift observations. We find that around multipoles $\ell\approx 100$, reionization induces $B$-mode power with $\ell(\ell+1)C_\ell^{BB}/2π\approx 4\times 10^{-6}\,μ$K$^2$. This secondary signal is thus at the level of the primordial signal with the tensor-to-scalar ratio $r<10^{-4}$, and can increase by a factor of $\sim 50$ if reionization is sourced by highly clustered sources residing in haloes with mass of $\sim 10^{11}$ M$_\odot$. Our findings suggest that the contribution of patchy reionization to the search for primordial gravitational waves is unlikely to be a concern for currently planned CMB experiments.

astro-ph.CO↗

arXiv2021

The correlation of high-redshift galaxies with the thermal Sunyaev-Zel'dovich effect traces reionization

We explore a potential new probe of reionization: the cross-correlation of high-redshift galaxies with maps of the thermal Sunyaev-Zel'dovich (tSZ) effect. We consider two types of high redshift galaxies: Lyman break galaxies (LBGs) and Lyman-$α$ emitters (LAEs). LBGs and LAEs will be detected in large numbers at high redshift ($z \approx$ 4 to 7) by ongoing and future surveys. We consider a LBG sample like that expected from the The Rubin Observatory Legacy Survey of Space and Time (LSST), and a selection of LAEs modelled after the Subaru SILVERRUSH program, but covering a much larger sky fraction. The tSZ effect is sensitive to a line-of-sight integral of the ionized gas pressure, and can be measured across large patches of the sky using multi-frequency CMB surveys. We consider forecast tSZ maps from CMB Stage 4 and more futuristic observations. Using a suite of hydrodynamical simulations, we show that the high-redshift galaxies are correlated with the tSZ signal from reionization, with cross-power amplitude of order $10^{-15}$ at $\ell \sim 1000$. The cross-spectra between LBGs/LAEs with tSZ maps contain information about the reionization history of the Universe, such as the distribution of bubble sizes, and could in principle be used to directly measure the timing of reionization. The amplitude of the signal is small, however, and its detectability is hindered by low-redshift contributions to tSZ maps and by instrumental noise. If the the low-redshift contribution to the observed tSZ signal is suppressed by masking of massive halos, a combination of overlapping futuristic CMB and galaxy surveys could potentially probe this signal.

astro-ph.CO↗

arXiv2020

Detecting Trojaned DNNs Using Counterfactual Attributions

We target the problem of detecting Trojans or backdoors in DNNs. Such models behave normally with typical inputs but produce specific incorrect predictions for inputs poisoned with a Trojan trigger. Our approach is based on a novel observation that the trigger behavior depends on a few ghost neurons that activate on trigger pattern and exhibit abnormally higher relative attribution for wrong decisions when activated. Further, these trigger neurons are also active on normal inputs of the target class. Thus, we use counterfactual attributions to localize these ghost neurons from clean inputs and then incrementally excite them to observe changes in the model's accuracy. We use this information for Trojan detection by using a deep set encoder that enables invariance to the number of model classes, architecture, etc. Our approach is implemented in the TrinityAI tool that exploits the synergies between trustworthiness, resilience, and interpretability challenges in deep learning. We evaluate our approach on benchmarks with high diversity in model architectures, triggers, etc. We show consistent gains (+10%) over state-of-the-art methods that rely on the susceptibility of the DNN to specific adversarial attacks, which in turn requires strong assumptions on the nature of the Trojan attack.

cs.LG↗

arXiv2020

Detectability of the $τ_{es}$-21cm cross-correlation: a tomographic probe of patchy reionization

The cross-correlation between fluctuations in the electron scattering optical depth $τ_{\rm es}$ as probed by future Cosmic Microwave Background (CMB) experiments, and fluctuations in the 21cm differential brightness temperature $ΔT_{\rm 21cm}$ as probed by ground-based radio interferometers, will trace the reionization history of the Universe. In particular, the $τ_{\rm es}-$21cm cross-correlation should yield a determination of the characteristic bubble size distribution and ionization fraction as a function of redshift. When assuming that the cross-correlation signal is limited by instrumental noise rather than by foregrounds, we estimate its potential detectability by upcoming experiments. Specifically, the combination of HERA and Simons Observatory, CMB-S4 and PICO should yield a signal-to-noise ratio around 3 - 6, while and the exploitation of the SKA should increase it to 10-20. Finally, we have discussed how such levels of detectability can be affected when (simply modeled) 21cm foregrounds are present. For the most promising PICO$\times$SKA configuration, an efficiency of foreground removal to a level of $7\times 10^{-4}$ is needed to achieve a $5σ$ detection of the cross-correlation signal; in addition, safe avoidance of foreground contamination in the line-of-sight Fourier modes above $0.03 \,h\,\rm Mpc^{-1}$ would guarantee a detection significance around $3σ$.

astro-ph.CO↗

arXiv2020

Beyond $Λ$CDM with Low and High Redshift Data: Implications for Dark Energy

Assuming that the Universe at higher redshifts (z \sim 4 and beyond) is consistent with LCDM model as constrained by the Planck measurements, we reanalyze the low redshift cosmological data to reconstruct the Hubble parameter as a function of redshift. This enables us to address the H_0 and other tensions between low z observations and high z Planck measurement from CMB. From the reconstructed H(z), we compute the energy density for the "dark energy" sector of the Universe as a function of redshift without assuming a specific model for dark energy. We find that the dark energy density has a minimum for certain redshift range and that the value of dark energy at this minimum is negative. This behavior can most simply be described by a negative cosmological constant plus an evolving dark energy component. We discuss possible theoretical and observational implications of such a scenario.

astro-ph.CO↗

arXiv2019

Polarimetric Radar Cross-Sections of Pedestrians at Automotive Radar Frequencies

Simulation of radar cross-sections (RCS) of pedestrians at automotive radar frequencies forms a key tool for software verification test beds for advanced driver assistance systems. Two commonly used simulation methods are: the computationally simple scattering center model of dynamic humans; and the shooting and bouncing ray technique based on geometric optics. The latter technique is more accurate but due to its computational complexity, it is usually used only for modeling scattered returns of still human poses. In this work, we combine the two methods in a linear regression framework to accurately estimate the scattering coefficients or reflectivies of point scatterers in a realistic automotive radar signal model which we subsequently use to simulate range-time, Doppler-time and range-Doppler radar signatures. The simulated signatures show a normalized mean square error below 10% and a structural similarity above $81\%$ with respect to measurement results generated with an automotive radar at 77 GHz.

eess.SP↗

arXiv2019

Align2Ground: Weakly Supervised Phrase Grounding Guided by Image-Caption Alignment

We address the problem of grounding free-form textual phrases by using weak supervision from image-caption pairs. We propose a novel end-to-end model that uses caption-to-image retrieval as a `downstream' task to guide the process of phrase localization. Our method, as a first step, infers the latent correspondences between regions-of-interest (RoIs) and phrases in the caption and creates a discriminative image representation using these matched RoIs. In a subsequent step, this (learned) representation is aligned with the caption. Our key contribution lies in building this `caption-conditioned' image encoding which tightly couples both the tasks and allows the weak supervision to effectively guide visual grounding. We provide an extensive empirical and qualitative analysis to investigate the different components of our proposed model and compare it with competitive baselines. For phrase localization, we report an improvement of 4.9% (absolute) over the prior state-of-the-art on the VisualGenome dataset. We also report results that are at par with the state-of-the-art on the downstream caption-to-image retrieval task on COCO and Flickr30k datasets.

cs.CV↗

arXiv2019

Cosmology With Low-Redshift Observations: No Signal For New Physics

We analyse various low-redshift cosmological data from Type-Ia Supernova, Baryon Acoustic Oscillations, Time-Delay measurements using Strong-Lensing, $H(z)$ measurements using Cosmic Chronometers and growth measurements from large scale structure observations for $Λ$CDM and some different dark energy models. By calculating the Bayesian Evidence for different dark energy models, we find out that the $Λ$CDM still gives the best fit to the data with $H_{0}=70.3^{+1.36}_{-1.35}$ Km/s/Mpc (at $1σ$). This value is in $2σ$ or less tension with various low and high redshift measurements for $H_{0}$ including SH0ES, Planck-2018 and the recent results from H0LiCOW-XIII. The derived constraint on $S_{8}=σ_{8}\sqrt{Ω_{m0}/{0.3}}$ from our analysis is $S_{8} = 0.76^{+0.03}_{-0.03}$, fully consistent with direct measurement of $S_{8}$ by KiDS+VIKING-450+DES1 survey. We hence conclude that the $Λ$CDM model with parameter constraints obtained in this work is consistent with different early and late Universe observations within $2σ$. We therefore, do not find any compelling reason to go beyond concordance $Λ$CDM model.

astro-ph.CO↗

arXiv2018

Understanding Visual Ads by Aligning Symbols and Objects using Co-Attention

We tackle the problem of understanding visual ads where given an ad image, our goal is to rank appropriate human generated statements describing the purpose of the ad. This problem is generally addressed by jointly embedding images and candidate statements to establish correspondence. Decoding a visual ad requires inference of both semantic and symbolic nuances referenced in an image and prior methods may fail to capture such associations especially with weakly annotated symbols. In order to create better embeddings, we leverage an attention mechanism to associate image proposals with symbols and thus effectively aggregate information from aligned multimodal representations. We propose a multihop co-attention mechanism that iteratively refines the attention map to ensure accurate attention estimation. Our attention based embedding model is learned end-to-end guided by a max-margin loss function. We show that our model outperforms other baselines on the benchmark Ad dataset and also show qualitative results to highlight the advantages of using multihop co-attention.

cs.CV↗

arXiv2018

Observing patchy reionization with future CMB polarization experiments

We study the signal from patchy reionization in view of the future high accuracy polarization measurements of the Cosmic Microwave Background (CMB). We implement an extraction procedure of the patchy reionization signal analogous to CMB lensing. We evaluate the signal to noise ratio (SNR) for the future Stage IV (S4) CMB experiment. The signal has a broad peak centered on the degree angular scales, with a long tail at higher multipoles. The CMB S4 experiment can effectively constrain the properties of reionization by measuring the signal on degree scales. The signal amplitude depends on the properties of the structure determining the reionization morphology. We describe bubbles having radii distributed log-normally. The expected S/N is sensitive to the mean bubble radius: $\bar{R}=5$ Mpc implies $S/N \approx 4$, $\bar{R}=10$ Mpc implies $S/N \approx 20$. The spread of the radii distribution strongly affects the integrated SNR, that changes by a factor of $10^2$ when $σ_{lnr}$ goes from $\ln 2$ to $\ln3$. Future CMB experiments will thus place important constraints on the physics of reionization.

astro-ph.CO↗

arXiv2011

Hardy's non-locality and generalized non-local theory

Hardy's non-locality theorem for multiple two-level systems is explored in the context of generalized nonlocal theory. We find nonlocal but non-signaling probabilities, providing Hardy's nonlocal argument, which are higher than those in Quantum Mechanics. Maximum probability of success of Hardy's argument is obtained for three two-level systems in quantum as well as in a more generalized theory. Interestingly, the maximum in the nonlocal generalized theory for both the cases turns out to be same.

quant-ph↗