A5056315092- Model Reduction and Neural Networks
- Adversarial Robustness in Machine Learning
- Generative Adversarial Networks and Image Synthesis
- Advanced Combustion Engine Technologies
- Combustion and flame dynamics
- Hydrocarbon exploration and reservoir analysis
- Advanced Neural Network Applications
- Anomaly Detection Techniques and Applications
- Cell Image Analysis Techniques
- Domain Adaptation and Few-Shot Learning
- Retinal Imaging and Analysis
- Rheology and Fluid Dynamics Studies
- Computational Physics and Python Applications
- Numerical methods in engineering
- Advanced Chemical Physics Studies
- Combustion and Detonation Processes
- Advanced Numerical Analysis Techniques
- Video Coding and Compression Technologies
- Advanced Memory and Neural Computing
- Imbalanced Data Classification Techniques
- Image and Video Quality Assessment
- Multimedia Communication and Technology
- Machine Learning in Materials Science
- Digital Media Forensic Detection
- Phase Equilibria and Thermodynamics
New York University
2020-2024
Iowa State University
2019-2020
Birla Institute of Technology and Science, Pilani - Goa Campus
2014
University of South Wales
2010
ExxonMobil (United States)
2005
University of Delaware
2002-2005
University of Southern California
2005
Deep neural networks have been shown to exhibit an intriguing vulnerability adversarial input images corrupted with imperceptible perturbations. However, the majority of attacks assume global, fine-grained control over image pixel space. In this paper, we consider a different setting: what happens if adversary could only alter specific attributes image? These would generate inputs that might be perceptibly different, but still natural-looking and enough fool classifier. We propose novel...
Abstract Cheap and ubiquitous sensing has made collecting large agricultural datasets relatively straightforward. These (for instance, citizen science data curation platforms like iNaturalist) can pave the way for developing powerful artificial intelligence (AI) models detection counting. However, traditional supervised learning methods require labeled data, manual annotation of these raw with useful labels (such as bounding boxes or segmentation masks) be extremely laborious, expensive,...
Abstract The rate coefficient of CO + OH → products is analyzed with RRKM/master equation analyses and Monte Carlo simulations. are based on the recent CCSD(T)/cc‐pvTZ potential energy surface Yu et al. Chem Phys Lett 2001, 349, 547–554). It shown that experimental data over temperature range 80–2500 K pressure from 1 Torr to 800 bar can be satisfactorily reproduced by lowering barrier for 2 H exit channel kcal/mol more importantly, considering an equilibrium factor in thermal constant...
Recognizing the activities causing distraction in real-world driving scenarios is critical for ensuring safety and reliability of both drivers pedestrians on roadways. Conventional computer vision techniques are typically data-intensive require a large volume annotated training data to detect classify various distracted behaviors, thereby limiting their generalization ability, efficiency scalability. We aim develop generalized framework that showcases robust performance with access limited...
The unimolecular decomposition of ethylene oxide (oxirane) and the oxiranyl radial is examined by molecular orbital calculations, Rice-Ramsperger-Kassel-Marcus (RRKM)/Master Equation analysis, detailed kinetic modeling pyrolysis in a single-pulse shock tube. It was found that largest energy barrier to lies its initial isomerization form acetaldehyde, agreement with previous studies, proceed through *CH2CH2O* biradical. Because biradical nature transition states intermediate, barriers for C-O...
Reliable training of generative adversarial networks (GANs) typically require massive datasets in order to model complicated distributions. However, several applications, samples obey invariances that are \textit{a priori} known; for example, complex physics simulations, the data universal laws encoded as well-defined mathematical equations. In this paper, we propose a new modeling approach, InvNet, can efficiently spaces with known invariances. We devise an algorithm encode them into...
We propose a two step framework to automatically classify an OCT scan as indicative of Diabetic Macular Edema (DME) by detecting abnormal pathologies in frames. The first involves detection candidate patches for fluid filled regions and hard exudates using image processing techniques. second is predict label these deep convolutional neural network. In the final collation step, we aggregate confidences CNN models use rule based method presence DME.
In this paper we propose a new family of algorithms, ATENT, for training adversarially robust deep neural networks. We formulate loss function that is equipped with an additional entropic regularization. Our considers the contribution adversarial samples are drawn from specially designed distribution in data space assigns high probability to points and immediate neighborhood samples. proposed algorithms optimize seek valleys landscape. approach achieves competitive (or better) performance...
Generative Adversarial Networks (GANs), while widely successful in modeling complex data distributions, have not yet been sufficiently leveraged scientific computing and design. Reasons for this include the lack of flexibility GANs to represent discrete-valued image data, as well control over physical properties generated samples. We propose a new conditional generative approach (InvNet) that efficiently enables images, allowing their parameterized geometric statistical properties. evaluate...
Vision transformers rely on a patch token based self attention mechanism, in contrast to convolutional networks. We investigate fundamental differences between these two families of models, by designing block sparsity adversarial attack. probe and analyze transformer as well models with attacks varying sizes. infer that are more sensitive than ResNets outperforming Transformer up $\sim30\%$ robust accuracy for single attacks.
Retinal Optical Coherence Tomography (OCT) scans are an important diagnostic tool for ophthalmologists. These provide a cross-sectional view of the retina ophthalmologists to detect abnormalities. A common type abnormality found in these is Fluid Filled Region (FFR). In this paper, we present method simultaneously classify and localize FFRs within retinal OCT using specialized Convolutional Neural Network (CNN). The training data weakly labeled, with only indication whether scan contains or...
The telecommunication industry has seen a major transition over the last decade. There been steady shift from voice to video communication which includes calls, television streaming and on demand services. introduction of Next Generation Networks led convergence fixed mobile devices. These devices have different capabilities such as screen sizes are connected through access networks support codecs. Due these differences it is impossible stream single format all In this paper we present real...
Deep neural networks are often highly overparameterized, prohibiting their use in compute-limited systems. However, a line of recent works has shown that the size deep can be considerably reduced by identifying subset neuron indicators (or mask) correspond to significant weights prior training. We demonstrate an simple iterative mask discovery method achieve state-of-the-art compression very networks. Our algorithm represents hybrid approach between single shot network pruning methods (such...
Recent progress in scientific machine learning (SciML) has opened up the possibility of training novel neural network architectures that solve complex partial differential equations (PDEs). Several (nearly data free) approaches have been recently reported successfully PDEs, with examples including deep feed forward networks, generative and encoder-decoder networks. However, practical adoption these is limited by difficulty models, especially to make predictions at large output resolutions (≥...