Resource
2019 EN
Behnam Gholami · Pritish Sahu · Minyoung Kim
+1 more
Domain Adaptation (DA), the process of effectively adapting task modelslearned on one domain, the source, to other related but distinct domains, thetargets, with no or minimal retraining, is typically accomplished using theprocess of source-to-target manifold alignment. However, this process oftenleads to unsatisfactory adaptation performance, in part because it ignores thetask-specific structure of the data. In this paper, we improve the performanceof DA by introducing a discriminative discrepancy measure which takes advantageof auxiliary information available in the source and the target domains tobetter align the source and target distributions. Specifically, we leverage thecohesive clustering structure within individual data manifolds, associated withdifferent tasks, to improve the alignment. This structure is explicit in thesource, where the task labels are available, but is implicit in the target,making the problem challenging. We address the challenge by devising a deep DAframework, which combines a new task-driven domain alignment discriminator withdomain regularizers that encourage the shared features as task-specific anddomain invariant, and prompt the task model to be data structure preserving,guiding its decision boundaries through the low density data regions. Wevalidate our framework on standard benchmarks, including Digits (MNIST, USPS,SVHN, MNIST-M), PACS, and VisDA. Our results show that our proposed modelconsistently outperforms the state-of-the-art in unsupervised domainadaptation.
Resource
2019 EN
Satya Narayan Shukla · Anit Kumar Sahu · Devin Willmott
+1 more
We focus on the problem of black-box adversarial attacks, where the aim is togenerate adversarial examples using information limited to loss functionevaluations of input-output pairs. We use Bayesian optimization~(BO) tospecifically cater to scenarios involving low query budgets to develop queryefficient adversarial attacks. We alleviate the issues surrounding BO inregards to optimizing high dimensional deep learning models by effectivedimension upsampling techniques. Our proposed approach achieves performancecomparable to the state of the art black-box adversarial attacks albeit with amuch lower average query count. In particular, in low query budget regimes, ourproposed method reduces the query count up to $80\%$ with respect to the stateof the art methods.
Resource
2019 EN
S. Sahoo · R. C. Baral · P. K. Sahu
+1 more
The mass dependence plays a significant role in the yield enhancement orsuppression of hadrons in pp and p-Pb collisions at the LHC energies. This hasbeen observed by parameterizing the variation of yield ratios between any twohadrons with event charged-particle multiplicity using a single empiricalfunction. We notice that this variation is independent of all quantum numbersand solely depends on masses of hadrons and masses of their valence quarks. Thefunction shows that the amount of quark deconfinement increases with eventmultiplicity, and the quark coalescence favours more the production of heavierhadrons compared to lighter ones.
Resource
2019 EN
Smruti R. Sahoo · Rudranarayan Khatua · Suryakanti Debata
+2 more
The anisotropic charge carrier mobilities of two phenancene series compoundssuch as dibenzo[a,c]picene (DBP) and tribenzo[a,c,k]tetraphene (TBT) isinvestigated based on the first-principle calculations and Marcus-Hush theory.The molecular packing patterns in organic crystal play an important role fordeterming the charge carrier mobility and hence the device efficienciesdesigned from the organic materials. Among the studied molecules, TBT shows amaximum anisotropic hole ($\mu_h=0.129\ cm^2V^{-1}s^{-1}$) and electron($\mu_h=1.834\ cm^2V^{-1}s^{-1}$) mobility, hence possesses an ambipolarsemiconducting character. The frontier molecular orbital analyses proved thebetter air-stability of the studied compounds than the conventional pentacene,because of their higher HOMO energy levels. Band structure calculations of thestudied compounds have also been investigated. From non-linear optical (NLO)properties anysis, we found the TBT compound shows more NLO response than DBP.
Resource
2019 EN
Divya Kaushik · Utkarsh Singh · Upasana Sahu
+2 more
Resistive Random Access Memory (RRAM) and Phase Change Memory (PCM) deviceshave been popularly used as synapses in crossbar array based analog NeuralNetwork (NN) circuit to achieve more energy and time efficient dataclassification compared to conventional computers. Here we demonstrate theadvantages of recently proposed spin orbit torque driven Domain Wall (DW)device as synapse compared to the RRAM and PCM devices with respect to on-chiplearning (training in hardware) in such NN. Synaptic characteristic of DWsynapse, obtained by us from micromagnetic modeling, turns out to be much morelinear and symmetric (between positive and negative update) than that of RRAMand PCM synapse. This makes design of peripheral analog circuits for on-chiplearning much easier in DW synapse based NN compared to that for RRAM and PCMsynapses. We next incorporate the DW synapse as a Verilog-A model in thecrossbar array based NN circuit we design on SPICE circuit simulator.Successful on-chip learning is demonstrated through SPICE simulations on thepopular Fisher's Iris dataset. Time and energy required for learning turn outto be orders of magnitude lower for DW synapse based NN circuit compared tothat for RRAM and PCM synapse based NN circuits.
Resource
2019 EN
Saurabh Sahu · Rahul Gupta · Carol Espy-Wilson
Emotion recognition is a classic field of research with a typical setupextracting features and feeding them through a classifier for prediction. Onthe other hand, generative models jointly capture the distributionalrelationship between emotions and the feature profiles. Relatively recently,Generative Adversarial Networks (GANs) have surfaced as a new class ofgenerative models and have shown considerable success in modeling distributionsin the fields of computer vision and natural language understanding. In thiswork, we experiment with variants of GAN architectures to generate featurevectors corresponding to an emotion in two ways: (i) A generator is trainedwith samples from a mixture prior. Each mixture component corresponds to anemotional class and can be sampled to generate features from the correspondingemotion. (ii) A one-hot vector corresponding to an emotion can be explicitlyused to generate the features. We perform analysis on such models and alsopropose different metrics used to measure the performance of the GAN models intheir ability to generate realistic synthetic samples. Apart from evaluation ona given dataset of interest, we perform a cross-corpus study where we study theutility of the synthetic samples as additional training data in low resourceconditions.
Resource
2019 EN
R. Baral · H. S. Fierro · C. Rueda
+11 more
Features of low dimensional magnetism resulting from a square-net arrangementof Co atoms in trirutile CoTa$_2$O$_6$ is studied in the present work by meansof density functional theory and is compared with the experimental results ofspecific heat and neutron diffraction. The small total energy differencesbetween the ferromagnetic (FM) and antiferromagnetic (AFM) configuration ofCoTa$_2$O$_6$ shows that competing magnetic ground states exist, with thepossibility of transition from FM to AFM phase at low temperature. Ourcalculation further suggests the semi-conducting behavior for CoTa$_2$O$_6$with a band gap of $\sim$0.41 eV. The calculated magnetic anisotropy energy is$\sim$2.5 meV with its easy axis along the [100] (in-plane) direction. Studyingthe evolution of magnetism in Co$_{1-x}$Mg$_x$Ta$_2$O$_6$ (x = 0, 0.1, 0.3,0.5, 0.7 and 1). it is found that the sharp AFM transition exhibited byCoTa$_2$O$_6$ at $T_N$ = 6.2 K in its heat capacity vanishes with Mg-dilution,indicating the obvious effect of weakening the superexchange pathways of Co.The current specific heat study reveals the robust nature of $T_N$ forCoTa$_2$O$_6$ in applied magnetic fields. Clear indication of short-rangemagnetism is obtained from the magnetic entropy, however, diffuse componentsare absent in neutron diffraction data. At $T_N$, CoTa$_2$O$_6$ enters along-range ordered magnetic state which can be described using a propagationvector, (1/4, 1/4, 0). Upon Mg-dilution at $x \geq$0.1, the long-range orderedmagnetism is destroyed. The present results should motivate an investigation ofmagnetic excitations in this low-dimensional anisotropic magnet.
Resource
2019 EN
Aritra Dutta · El Houcine Bergou · Ahmed M. Abdelmoniem
+4 more
Compressed communication, in the form of sparsification or quantization ofstochastic gradients, is employed to reduce communication costs in distributeddata-parallel training of deep neural networks. However, there exists adiscrepancy between theory and practice: while theoretical analysis of mostexisting compression methods assumes compression is applied to the gradients ofthe entire model, many practical implementations operate individually on thegradients of each layer of the model. In this paper, we prove that layer-wisecompression is, in theory, better, because the convergence rate is upperbounded by that of entire-model compression for a wide range of biased andunbiased compression methods. However, despite the theoretical bound, ourexperimental study of six well-known methods shows that convergence, inpractice, may or may not be better, depending on the actual trained model andcompression ratio. Our findings suggest that it would be advantageous for deeplearning frameworks to include support for both layer-wise and entire-modelcompression.
Resource
2019 EN
Puja Sahu · Nandyala Hemachandra
PAC-Bayesian set up involves a stochastic classifier characterized by aposterior distribution on a classifier set, offers a high probability bound onits averaged true risk and is robust to the training sample used. For a givenposterior, this bound captures the trade off between averaged empirical riskand KL-divergence based model complexity term. Our goal is to identify anoptimal posterior with the least PAC-Bayesian bound. We consider a finiteclassifier set and 5 distance functions: KL-divergence, its Pinsker's and asixth degree polynomial approximations; linear and squared distances. Lineardistance based model results in a convex optimization problem. We obtain closedform expression for its optimal posterior. For uniform prior, this posteriorhas full support with weights negative-exponentially proportional to number ofmisclassifications. Squared distance and Pinsker's approximation bounds arepossibly quasi-convex and are observed to have single local minimum. We derivefixed point equations (FPEs) using partial KKT system with strict positivityconstraints. This obviates the combinatorial search for subset support of theoptimal posterior. For uniform prior, exponential search on a full-dimensionalsimplex can be limited to an ordered subset of classifiers with increasingempirical risk values. These FPEs converge rapidly to a stationary point, evenfor a large classifier set when a solver fails. We apply these approaches toSVMs generated using a finite set of SVM regularization parameter values on 9UCI datasets. These posteriors yield stochastic SVM classifiers with tightbounds. KL-divergence based bound is the tightest, but is computationallyexpensive due to non-convexity and multiple calls to a root finding algorithm.Optimal posteriors for all 5 distance functions have lowest 10% test errorvalues on most datasets, with linear distance being the easiest to obtain.
Resource
2019 EN
Narayanasami Sathyamurthy · Michael Baer · Satyam Ravi
+3 more
In the early universe, following the nucleosynthesis, conditions were rightfor recombination processes to take place yielding neutral atoms H, He and Li.The understanding so far in astrophysics is that the first molecule to beformed was HeH+ by radiative association (He + H+ -> HeH+ + h(nu) and He+ + H-> HeH+ + h(nu). The recent report by Guesten et al (Nature, 568, 357, 2019) ofdetection of HeH+ in planetary Nebula NGC 7027 confirms its presence, but itdoes not conclusively prove the origin of this species. To create moleculesfrom free moving quasi-ions surrounded by an electronic cloud, theBorn-Oppenheimer-Huang (BOH) theory furnishes two kinds of forces, namely, onethat results from the Potential Energy Surfaces (PESs) and the other fromNon-Adiabatic Coupling Terms (NACTs). Whereas the PESs are known to manage slowmoving quasi-ions the NACTs, with their, frequently, infinitely large values atthe vicinity of the singularities can control the fast moving quasi-ions. Toachieve that the BOH equation indicates that the NACTs are affecting the fastmoving quasi-ions directly and if they are attributed with dissipative featuresor in other words to behave as a Friction Force they indeed could serve (likeany other ordinary friction) as moderators for the fast atomic(ionic) species.It is proposed in the present paper that the triatomic HeH2+ was the precursorto HeH+ and it could have been formed by the (He, H, H)+ nuclei coming togetherunder the electron cloud, facilitated by the NACTs between different electronicstates acting as an astronomical friction force. This is possible because ofthe singularities in the NACTs for triatomic systems and NOT for diatomicsystems. Although the existence of HeH2+ was established in the laboratory in1996, it has not been detected in the interstellar media so far. But, there isno reason why it cannot be detected in near future.