arXiv subjects

Anushree Ghosh

Publications and source records attributed to Anushree Ghosh.

13 recordsLinked to original sources

arXiv2026

MPSelectTune: Prompt-type Selection for Fine-tuning improves Concept Unlearning in LLMs

LLMs can be conveniently adapted to a diverse set of tasks, e.g, prediction, question-answering tasks, etc, using appropriate prompts with few-shot examples. Biased or harmful concepts, e.g. gender or bio-weapons, present in pre-trained LLMs can lead to unsafe or unethical responses for many such prompts. Removing such undesirable concepts robustly across different prompt types remains a challenging problem, since existing unlearning methods typically ignore the impact of prompt variation. In this paper, we explore a novel adversarial approach to use a joint prompt for the main task and concept task prediction. We show that fine-tuning using the ``worst prompt type'' for concept prediction (with the highest concept accuracy) improves the average unlearning performance over a fine-tuning method that uses a combination of all prompt types. Our proposed method, MPSelectTune, is a two-stage approach that minimizes the concept accuracy of the highest accuracy-prompt type, after fine-tuning using a novel multi-task loss using multiple prompt types. Experimental results on four benchmarks show $2 - 15\%$ main task accuracy improvements over recent baselines and while reducing the worst-case concept accuracy by up to $17\%$ compared to recent baselines.

cs.LG↗

arXiv2020

Inferring Convolutional Neural Networks' accuracies from their architectural characterizations

Convolutional Neural Networks (CNNs) have shown strong promise for analyzing scientific data from many domains including particle imaging detectors. However, the challenge of choosing the appropriate network architecture (depth, kernel shapes, activation functions, etc.) for specific applications and different data sets is still poorly understood. In this paper, we study the relationships between a CNN's architecture and its performance by proposing a systematic language that is useful for comparison between different CNN's architectures before training time. We characterize CNN's architecture by different attributes, and demonstrate that the attributes can be predictive of the networks' performance in two specific computer vision-based physics problems -- event vertex finding and hadron multiplicity classification in the MINERvA experiment at Fermi National Accelerator Laboratory. In doing so, we extract several architectural attributes from optimized networks' architecture for the physics problems, which are outputs of a model selection algorithm called Multi-node Evolutionary Neural Networks for Deep Learning (MENNDL). We use machine learning models to predict whether a network can perform better than a certain threshold accuracy before training. The models perform 16-20% better than random guessing. Additionally, we found an coefficient of determination of 0.966 for an Ordinary Least Squares model in a regression on accuracy over a large population of networks.

cs.CV↗

arXiv2018

Prospects of indirect searches for dark matter annihilations in the earth with ICAL@INO

We study the prospects of detecting muon events at the upcoming Iron CALorimeter (ICAL) detector to be built at the proposed India-based Neutrino Observatory (INO) facility due to neutrinos arising out of annihilation of Weakly Interactive Massive Particles (WIMP) in the centre of the earth. The atmospheric neutrinos coming from the direction of earth core presents an irreducible background. We consider 50kt $\times$ 10 years of ICAL running and WIMP masses between 10-100 GeV and present 90 \% C.L. exclusion sensitivity limits on $σ_{SI}$ which is the WIMP-nucleon Spin Independent (SI) interaction cross-section. The expected sensitivity limits calculated for ICAL for the WIMP annihilation in the earth are more stringent than the limits obtained by any other indirect detection experiment. For a WIMP mass of ~$52.14 \textup{ GeV}$, where the signal fluxes are enhanced due to resonance capture of WIMP in earth due to Fe nuclei, the sensitivity limits, assuming 100\% branching ratio for each channel, are : $σ_{SI} =1.02\times 10^{-44}~cm^2$ for the $τ^{+} τ^{-}$ channel and $σ_{SI} =5.36\times 10^{-44} ~cm^2$ for the $b ~\bar{b}$ channel.

hep-ph↗

arXiv2018

Prospects of Indirect Searches for Dark Matter at INO

The annihilation of Weakly Interactive Massive Particles (WIMP) in the centre of the sun could give rise to neutrino fluxes. We study the prospects of searching for these neutrinos at the upcoming Iron CALorimeter (ICAL) detector to be housed at the India-based Neutrino Observatory (INO). We perform ICAL simulations to obtain the detector efficiencies and resolutions in order to simulate muon events in ICAL due to neutrinos coming from annihilation of WIMP in the mass range $m_χ= (3-100)$ GeV. The atmospheric neutrinos pose a major background for these indirect detection studies and can be reduced using the fact that the signal comes only from the direction of the sun. For a given WIMP mass, we find the opening angle $θ_{90}$ such that 90 \% of the signal events are contained within this angle and use this cone-cut criteria to reduce the atmospheric neutrino background. The reduced background is then weighted by the solar exposure function at INO to obtain the final background spectrum for a given WIMP mass. We perform a $χ^2$ analysis and present expected exclusion regions in the $σ_{SD}-m_χ$ and $σ_{SI}-m_χ$, where $σ_{SD}$ and $σ_{SI}$ are the WIMP-nucleon Spin-Dependent (SD) and Spin-Independent (SI) scattering cross-section, respectively. For a 10 years exposure and $m_χ=25$ GeV, the expected 90 \% C.L. exclusion limit is found to be $σ_{SD} < 6.87\times 10^{-41}$ cm$^2$ and $σ_{SI} < 7.75\times 10^{-43}$ cm$^2$ for the $τ^{+} τ^{-}$ annihilation channel and $σ_{SD} < 1.14\times 10^{-39}$ cm$^2$ and $σ_{SI} < 1.30\times 10^{-41}$ cm$^2$ for the $b~\bar b $ channel, assuming 100 \% branching ratio for each of the WIMP annihilation channel.

hep-ex↗

arXiv2017

Search for the sterile neutrino mixing with the ICAL detector at INO

The study has been carried out on the prospects of probing the sterile neutrino mixing with the magnetized Iron CALorimeter (ICAL) at the India-based Neutrino Observatory (INO), using atmospheric neutrinos as a source. The so-called 3~$+$~1 scenario is considered for active-sterile neutrino mixing and lead to projected exclusion curves in the sterile neutrino mass and mixing angle plane. The analysis is performed using the neutrino event generator NUANCE, modified for ICAL, and folded with the detector resolutions obtained by the INO collaboration from a full GEANT4 based detector simulation. A comparison has been made between the results obtained from the analysis considering only the energy and zenith angle of the muon and combined with the hadron energy due to the neutrino induced event. A small improvement has been observed with the addition of the hadron information to the muon. In the analysis we consider neutrinos coming from all zenith angles and the Earth matter effects are also included. The inclusion of events from all zenith angles improves the sensitivity to sterile neutrino mixing by about 35$\%$ over the result obtained using only down-going events. The improvement mainly stems from the impact of Earth matter effects on active-sterile mixing. The expected precision of ICAL on the active-sterile mixing is explored and allowed confidence level (C.L.) contours presented. At the assumed true value of $10^\circ$ for the sterile mixing angles and marginalization over $Δm^2_{41}$ and the sterile mixing angles, the upper bound at 90\% C.L. (from 2 parameter plots) is around $20^\circ$ for $θ_{14}$ and $θ_{34}$, and about $12^\circ$ for $θ_{24}$.

hep-ph↗

arXiv2017

Physics Potential of the ICAL detector at the India-based Neutrino Observatory (INO)

The upcoming 50 kt magnetized iron calorimeter (ICAL) detector at the India-based Neutrino Observatory (INO) is designed to study the atmospheric neutrinos and antineutrinos separately over a wide range of energies and path lengths. The primary focus of this experiment is to explore the Earth matter effects by observing the energy and zenith angle dependence of the atmospheric neutrinos in the multi-GeV range. This study will be crucial to address some of the outstanding issues in neutrino oscillation physics, including the fundamental issue of neutrino mass hierarchy. In this document, we present the physics potential of the detector as obtained from realistic detector simulations. We describe the simulation framework, the neutrino interactions in the detector, and the expected response of the detector to particles traversing it. The ICAL detector can determine the energy and direction of the muons to a high precision, and in addition, its sensitivity to multi-GeV hadrons increases its physics reach substantially. Its charge identification capability, and hence its ability to distinguish neutrinos from antineutrinos, makes it an efficient detector for determining the neutrino mass hierarchy. In this report, we outline the analyses carried out for the determination of neutrino mass hierarchy and precision measurements of atmospheric neutrino mixing parameters at ICAL, and give the expected physics reach of the detector with 10 years of runtime. We also explore the potential of ICAL for probing new physics scenarios like CPT violation and the presence of magnetic monopoles.

physics.ins-det↗

arXiv2015

Neutrino Physics with Non-Standard Interactions at INO

Non-standard neutrino interactions (NSI) involved in neutrino propagation inside Earth matter could potentially alter atmospheric neutrino fluxes. In this work, we look at the impact of these NSI on the signal at the ICAL detector to be built at the India-based Neutrino Observatory (INO). We show how the sensitivity to the neutrino mass hierarchy of ICAL changes in the presence of NSI. The mass hierarchy sensitivity is shown to be rather sensitive to the NSI parameters $ε_{eμ}$ and $ε_{eτ}$, while the dependence on $ε_{μτ}$ and $ε_{ττ}$ is seen to be very mild, once the $χ^2$ is marginalised over oscillation and NSI parameters. If the NSI are large enough, the event spectrum at ICAL is expected to be altered and this can be used to discover new physics. We calculate the lower limit on NSI parameters above which ICAL could discover NSI at a given C.L. from 10 years of data. If NSI were too small, the null signal at ICAL can constrain the NSI parameters. We give upper limits on the NSI parameters at any given C.L. that one is expected to put from 10 years of running of ICAL. Finally, we give C.L. contours in the NSI parameter space that is expected to be still allowed from 10 years of running of the experiment.

hep-ph↗

arXiv2015

Simulation studies of hadron energy resolution as a function of iron plate thickness at INO-ICAL

We report on a detailed simulation study of the hadron energy resolution as a function of the thickness of the absorber plates for the proposed Iron Calorimeter (ICAL) detector at the India-based Neutrino Observatory (INO). We compare the hadron resolutions obtained with absorber thicknesses in the range 1.5--8 cm for neutrino interactions in the energy range 2--15 GeV, which is relevant to hadron production in atmospheric neutrino interactions. We find that at lower energies, the thickness dependence of energy resolution is steeper than at higher energies, however there is a thickness-independent contribution that dominates at the lower thicknesses discussed in this work. As a result, the gain in hadron energy resolution with decreasing plate thickness is marginal. We present the results in the form of fits to a function with energy-dependent exponent.

physics.ins-det↗

arXiv2013

Hadron energy response of the Iron Calorimeter detector at the India-based Neutrino Observatory

The results of a Monte Carlo simulation study of the hadron energy response for the magnetized Iron CALorimeter detector, ICAL, proposed to be located at the India-based Neutrino Observatory (INO) is presented. Using a GEANT4 modeling of the detector ICAL, interactions of atmospheric neutrinos with target nuclei are simulated. The detector response to hadrons propagating through it is investigated using the hadron hit multiplicity in the active detector elements. The detector response to charged pions of fixed energy is studied first, followed by the average response to the hadrons produced in atmospheric neutrino interactions using events simulated with the NUANCE event generator. The shape of the hit distribution is observed to fit the Vavilov distribution, which reduces to a Gaussian at high energies. In terms of the parameters of this distribution, we present the hadron energy resolution as a function of hadron energy, and the calibration of hadron energy as a function of the hit multiplicity. The energy resolution for hadrons is found to be in the range 85% (for 1GeV) -- 36% (for 15 GeV).

physics.ins-det↗

arXiv2013

Determining the Octant of $θ_{23}$ with PINGU, T2K, NOvA and Reactor Data

We explore the prospects of determining the octant of $θ_{23}$ with atmospheric neutrinos at PINGU. We study in detail the impact of energy and angle resolutions of the neutrino on the octant sensitivity. We show that the systematic uncertainties on the atmospheric neutrino flux predictions, especially the ones which affect the energy and zenith angle spectrum of the neutrinos, make a rather drastic reduction of the sensitivity of PINGU. We also study the prospects of measuring the octant of $θ_{23}$ in the long baseline experiments T2K and NOvA in conjunction with the reactor experiments. We study this for two configurations of NOvA and T2K and make a comparative analysis of them. From just 3 years of PINGU data, the octant could be determined at more than $3σ$ C.L. for $sin^2θ_{23}<0.419$ and $sin^2θ_{23}>0.586$ if we add the reactor data and if normal hierarchy is true. On addition of the data from T2K and NOvA, the sensitivity improves so that the octant could be determined at the $4σ$ C.L. for $sin^2θ_{23}<0.426$ and $sin^2θ_{23}>0.586$ if normal hierarchy is true. Even a $5σ$ significance for the right octant can be achieved if $sin^2θ_{23}<0.413$ for the true normal hierarchy. The sensitivity for the true inverted hierarchy is lower and we expect a $3σ$ determination of octant for $sin^2θ_{23}<0.43$ and $>0.585$ from the combined data set for this case.

hep-ph↗

arXiv2013

The Reach of INO for Atmospheric Neutrino Oscillation Parameters

The India-based Neutrino Observatory (INO) will host a 50 kt magnetized iron calorimeter (ICAL@INO) for the study of atmospheric neutrinos. Using the detector resolutions and efficiencies obtained by the INO collaboration from a full-detector GEANT4-based simulation, we determine the reach of this experiment for the measurement of the atmospheric neutrino mixing parameters ($\sin^2 θ_{23}$ and $|Δm_{32}^2 |$). We also explore the sensitivity of this experiment to the deviation of $θ_{23}$ from maximal mixing, and its octant.

hep-ph↗

arXiv2013

Measuring the Mass Hierarchy with Muon and Hadron Events in Atmospheric Neutrino Experiments

Neutrino mass hierarchy can be measured in atmospheric neutrino experiments through the observation of earth matter effects. Magnetized iron calorimeters have been shown to be good in this regard due to their charge identification capabilities. The charged current interaction of $ν_μ$ in this detector, produces a muon track and a hadron shower. The direction of the muon track can be measured very accurately. We show the improvement expected in the reach of this class of experiments to the neutrino mass hierarchy, as we improve the muon energy resolution and the muon reconstruction efficiency. We next propose to include the hadron events in the analysis, by tagging them with the zenith angle of the corresponding muon and binning the hadron data first in energy and then in zenith angle. To the best of our knowledge this way of performing the analysis of the atmospheric neutrino data has not be considered before. We show that the hadron events increase the mass hierarchy sensitivity of the experiment. Finally, we show the expected mass hierarchy sensitivity in terms of the reconstructed neutrino energy and zenith angle. We show how the detector resolutions spoil the earth matter effects in the neutrino channel and argue why the sensitivity obtained from the neutrino analysis cannot be significantly better than that obtained from the analysis using muon data alone. As a result, the best mass hierarchy sensitivity is obtained when we add the contribution of the muon and the hadron data. For $\sin^22θ_{13}=0.1$, $\sin^2θ_{23}=0.5$, a muon energy resolution of 2\%, reconstruction efficiency of 80\% and exposure of $50\times 10$ kton-year, we could get up to $4.5σ$ signal for the mass hierarchy from combining the muon and hadron data.

hep-ph↗

arXiv2013

Determining the Neutrino Mass Hierarchy with INO, T2K, NOvA and Reactor Experiments

The relatively large measured value of $θ_{13}$ has opened up the possibility of determining the neutrino mass hierarchy through earth matter effects. Amongst the current accelerator-based experiments only NOvA has a long enough baseline to observe earth matter effects. However, NOvA is plagued with uncertainty on the knowledge of the true value of $δ_{CP}$, and this could drastically reduce its sensitivity to the neutrino mass hierarchy. The earth matter effect on atmospheric neutrinos on the other hand is almost independent of $δ_{CP}$. The 50 kton magnetized Iron CALorimeter at the India-based Neutrino Observatory (ICAL@INO) will be observing atmospheric neutrinos. The charge identification capability of this detector gives it an edge over others for mass hierarchy determination through observation of earth matter effects. We study in detail the neutrino mass hierarchy sensitivity of the data from this experiment simulated using the Nuance based generator developed for ICAL@INO and folded with the detector resolutions and efficiencies obtained by the INO collaboration from a full Geant4-based detector simulation. The data from ICAL@INO is then combined with simulated data from T2K, NOvA, Double Chooz, RENO and Daya Bay experiments and a combined sensitivity study to the mass hierarchy is performed. With 10 years of ICAL@INO data combined with T2K, NOvA and reactor data, one could get about $2.3σ-5.7σ$ discovery of the neutrino mass hierarchy, depending on the true value of $\sin^2θ_{23}$ [0.4 -- 0.6], $\sin^22θ_{13}$ [0.08 -- 0.12] and $δ_{CP}$ [0 -- 2$π$].

hep-ph↗