Hubs
Live
My Hubs
All
My Hubs
All

### Trending Users

Emeka A. Ezewudo
Prateek Jassal
Patrick Joyce
Eric Cuellar
Titus Osikhiana Ogahbrai
Akhil Kunche
Ramees P S
dhanu dhanu
Sebastian Hunte

Trending
Today
Trending
Today

# Sign in to discover all of the research papers you care about, live as they're published.

10
Authors: Lyu, Siwei, et al
Published: Mar 2021
Authors: Lyu, Siwei, et al
Published: Mar 2021
In recent years, the advent of deep learning-based techniques and thesignificant reduction in the cost of computation resulted in the feasibility ofcreating realistic videos of human faces, commonly known as DeepFakes. Theavailability of open-source tools to create DeepFakes poses as a threat to thetrustworthiness of the online media. In this work, we develop an open-sourceonline platform, known as DeepFake-o-meter, that integrates state-of-the-artDeepFake detection methods and provide a convenient interface for the users. Wedescribe the design and function of DeepFake-o-meter in this work.
Slide 1 of 1
1
Authors: Gidel, Gauthier, et al
Published: Mar 2021
Authors: Gidel, Gauthier, et al
Published: Mar 2021
Adversarial attacks expose important vulnerabilities of deep learning models,yet little attention has been paid to settings where data arrives as a stream.In this paper, we formalize the online adversarial attack problem, emphasizingtwo key elements found in real-world use-cases: attackers must operate underpartial knowledge of the target model, and the decisions made by the attackerare irrevocable since they operate on a transient data stream. We firstrigorously analyze a deterministic variant of the online threat model bydrawing parallels to the well-studied $k$-\textit{secretary problem} andpropose \algoname, a simple yet practical algorithm yielding a provably bettercompetitive ratio for $k=2$ over the current best single threshold algorithm.We also introduce the \textit{stochastic $k$-secretary} -- effectively reducingonline blackbox attacks to a $k$-secretary problem under noise -- and provetheoretical bounds on the competitive ratios of \textit{any} online algorithmsadapted to this setting. Finally, we complement our theoretical results byconducting a systematic suite of experiments on MNIST and CIFAR-10 with bothvanilla and robust classifiers, revealing that, by leveraging online secretaryalgorithms, like \algoname, we can get an online attack success rate close tothe one achieved by the optimal offline solution.
Slide 1 of 1
1
Authors: Kaufman, Mike, et al
Published: Mar 2021
Authors: Kaufman, Mike, et al
Published: Mar 2021
Forking structure is widespread in the open-source repositories and thatcauses a significant number of merge conflicts. In this paper, we study theproblem of textual merge conflicts from the perspective of Microsoft Edge, alarge, highly collaborative fork off the main Chromium branch with significantmerge conflicts. Broadly, this study is divided into two sections. First, weempirically evaluate textual merge conflicts in Microsoft Edge and classifythem based on the type of files, location of conflicts in a file, and the sizeof conflicts. We found that ~28% of the merge conflicts are 1-2 line changes,and many resolutions have frequent patterns. Second, driven by these findings,we explore Program Synthesis (for the first time) to learn patterns and resolvestructural merge conflicts. We propose a novel domain-specific language (DSL)that captures many of the repetitive merge conflict resolution patterns andlearn resolution strategies as programs in this DSL from example resolutions.We found that the learned strategies can resolve 11.4% of the conflicts (~41%of 1-2 line changes) that arise in the C++ files with 93.2% accuracy.
Slide 1 of 1
1
Authors: Seán McLoone, Federico Zocco, Seán McLoone
Published: Mar 2021
Authors: Seán McLoone, Federico Zocco, Seán McLoone
Published: Mar 2021
Long-term availability of minerals and industrial materials is a necessarycondition for sustainable development as they are the constituents of anymanufacturing product. In particular, technologies with increasing demand suchas GPUs and photovoltaic panels are made of critical raw materials. To enhancethe efficiency of material management, in this paper we make three maincontributions: first, we identify in the literature an emergingcomputer-vision-enabled material monitoring technology which we call MaterialMeasurement Unit (MMU); second, we provide a survey of works relevant to thedevelopment of MMUs; third, we describe a material stock monitoring sensornetwork deploying multiple MMUs.
Slide 1 of 1
1
Authors: Ehsan, Shoaib, et al
Published: Mar 2021
Authors: Ehsan, Shoaib, et al
Published: Mar 2021
Visual Place Recognition (VPR) is the ability to correctly recall apreviously visited place using visual information under environmental,viewpoint and appearance changes. An emerging trend in VPR is the use ofsequence-based filtering methods on top of single-frame-based place matchingtechniques for route-based navigation. The combination leads to varying levelsof potential place matching performance boosts at increased computationalcosts. This raises a number of interesting research questions: How doesperformance boost (due to sequential filtering) vary along the entire spectrumof single-frame-based matching methods? How does sequence matching lengthaffect the performance curve? Which specific combinations provide a goodtrade-off between performance and computation? However, there is lack ofprevious work looking at these important questions and most of thesequence-based filtering work to date has been used without a systematicapproach. To bridge this research gap, this paper conducts an in-depthinvestigation of the relationship between the performance of single-frame-basedplace matching techniques and the use of sequence-based filtering on top ofthose methods. It analyzes individual trade-offs, properties and limitationsfor different combinations of single-frame-based and sequential techniques. Anumber of state-of-the-art VPR methods and widely used public datasets areutilized to present the findings that contain a number of meaningful insightsfor the VPR community.
Slide 1 of 1
1
Authors: Faust, Aleksandra, et al
Published: Mar 2021
Authors: Faust, Aleksandra, et al
Published: Mar 2021
Slide 1 of 1
1
Authors: Bojanowski, Piotr, et al
Published: Mar 2021
Authors: Bojanowski, Piotr, et al
Published: Mar 2021
Recently, self-supervised learning methods like MoCo, SimCLR, BYOL and SwAVhave reduced the gap with supervised methods. These results have been achievedin a control environment, that is the highly curated ImageNet dataset. However,the premise of self-supervised learning is that it can learn from any randomimage and from any unbounded dataset. In this work, we explore ifself-supervision lives to its expectation by training large models on random,uncurated images with no supervision. Our final SElf-supERvised (SEER) model, aRegNetY with 1.3B parameters trained on 1B random images with 512 GPUs achieves84.2% top-1 accuracy, surpassing the best self-supervised pretrained model by1% and confirming that self-supervised learning works in a real world setting.Interestingly, we also observe that self-supervised models are good few-shotlearners achieving 77.9% top-1 with access to only 10% of ImageNet. Code:https://github.com/facebookresearch/vissl
Slide 1 of 1
1
Published: Mar 2021
Published: Mar 2021
AI application developers typically begin with a dataset of interest and avision of the end analytic or insight they wish to gain from the data at hand.Although these are two very important components of an AI workflow, one oftenspends the first few weeks (sometimes months) in the phase we refer to as dataconditioning. This step typically includes tasks such as figuring out how toprepare data for analytics, dealing with inconsistencies in the dataset, anddetermining which algorithm (or set of algorithms) will be best suited for theapplication. Larger, faster, and messier datasets such as those from Internetof Things sensors, medical devices or autonomous vehicles only amplify theseissues. These challenges, often referred to as the three Vs (volume, velocity,variety) of Big Data, require low-level tools for data management, preparationand integration. In most applications, data can come from structured and/orunstructured sources and often includes inconsistencies, formattingdifferences, and a lack of ground-truth labels. In this report, we highlight a number of tools that can be used to simplifydata integration and preparation steps. Specifically, we focus on dataintegration tools and techniques, a deep dive into an exemplar data integrationtool, and a deep-dive in the evolving field of knowledge graphs. Finally, weprovide readers with a list of practical steps and considerations that they canuse to simplify the data integration challenge. The goal of this report is toprovide readers with a view of state-of-the-art as well as practical tips thatcan be used by data creators that make data integration more seamless.
Slide 1 of 1
1
Authors: Michopoulos, John, et al
Published: Mar 2021
Authors: Michopoulos, John, et al
Published: Mar 2021
This work presents a data-driven reduced-order modeling framework toaccelerate the computations of $N$-body dynamical systems and their pair-wiseinteractions. The proposed framework differs from traditional accelerationmethods, like the Barnes-Hut method, which requires online tree building of thestate space, or the fast-multipole method, which requires rigorous $a$ $priori$analysis of governing kernels and online tree building. Our approach combinesBarnes-Hut hierarchical decomposition, dimensional compression via theleast-squares Petrov-Galerkin (LSPG) projection, and hyper-reduction by way ofthe Gauss-Newton with approximated tensor (GNAT) approach. The resulting$projection-tree$ reduced order model (PTROM) enables a drastic reduction inoperational count complexity by constructing sparse hyper-reduced pairwiseinteractions of the $N$-body dynamical system. As a result, the presentedframework is capable of achieving an operational count complexity that isindependent of $N$, the number of bodies in the numerical domain. Capabilitiesof the PTROM method are demonstrated on the two-dimensional fluid-dynamicBiot-Savart kernel within a parametric and reproductive setting. Results showthe PTROM is capable of achieving over 2000$\times$ wall-time speed-up withrespect to the full-order model, where the speed-up increases with $N$. Theresulting solution delivers quantities of interest with errors that are lessthan 0.1$\%$ with respect to full-order model.
Slide 1 of 1
1
Authors: Savir, Yonatan, et al
Published: Mar 2021
Authors: Savir, Yonatan, et al
Published: Mar 2021
Background. Eosinophilic esophagitis (EoE) is an allergic inflammatorycondition of the esophagus associated with elevated numbers of eosinophils.Disease diagnosis and monitoring requires determining the concentration ofeosinophils in esophageal biopsies, a time-consuming, tedious and somewhatsubjective task currently performed by pathologists. Methods. Herein, we aimedto use machine learning to identify, quantitate and diagnose EoE. We labeledmore than 100M pixels of 4345 images obtained by scanning whole slides ofH&E-stained sections of esophageal biopsies derived from 23 EoE patients. Weused this dataset to train a multi-label segmentation deep network. To validatethe network, we examined a replication cohort of 1089 whole slide images from419 patients derived from multiple institutions. Findings. PECNet segmentedboth intact and not-intact eosinophils with a mean intersection over union(mIoU) of 0.93. This segmentation was able to quantitate intact eosinophilswith a mean absolute error of 0.611 eosinophils and classify EoE diseaseactivity with an accuracy of 98.5%. Using whole slide images from thevalidation cohort, PECNet achieved an accuracy of 94.8%, sensitivity of 94.3%,and specificity of 95.14% in reporting EoE disease activity. Interpretation. Wehave developed a deep learning multi-label semantic segmentation network thatsuccessfully addresses two of the main challenges in EoE diagnostics anddigital pathology, the need to detect several types of small featuressimultaneously and the ability to analyze whole slides efficiently. Our resultspave the way for an automated diagnosis of EoE and can be utilized for otherconditions with similar challenges.
Slide 1 of 1