arXiv subjects

Amir Bar

Publications and source records attributed to Amir Bar.

49 records · Page 3Linked to original sources

arXiv2020

Learning Canonical Representations for Scene Graph to Image Generation

Generating realistic images of complex visual scenes becomes challenging when one wishes to control the structure of the generated images. Previous approaches showed that scenes with few entities can be controlled using scene graphs, but this approach struggles as the complexity of the graph (the number of objects and edges) increases. In this work, we show that one limitation of current methods is their inability to capture semantic equivalence in graphs. We present a novel model that addresses these issues by learning canonical graph representations from the data, resulting in improved image generation for complex visual scenes. Our model demonstrates improved empirical performance on large scene graphs, robustness to noise in the input scene graph, and generalization on semantically equivalent graphs. Finally, we show improved performance of the model on three different benchmarks: Visual Genome, COCO, and CLEVR.

cs.CV↗

arXiv2019

Improved ICH classification using task-dependent learning

Head CT is one of the most commonly performed imaging studied in the Emergency Department setting and Intracranial hemorrhage (ICH) is among the most critical and timesensitive findings to be detected on Head CT. We present BloodNet, a deep learning architecture designed for optimal triaging of Head CTs, with the goal of decreasing the time from CT acquisition to accurate ICH detection. The BloodNet architecture incorporates dependency between the otherwise independent tasks of segmentation and classification, achieving improved classification results. AUCs of 0.9493 and 0.9566 are reported on held out positive-enriched and randomly sampled sets comprised of over 1400 studies acquired from over 10 different hospitals. These results are comparable to previously reported results with smaller number of tagged studies.

cs.CV↗

arXiv2019

Learning Individual Styles of Conversational Gesture

Human speech is often accompanied by hand and arm gestures. Given audio speech input, we generate plausible gestures to go along with the sound. Specifically, we perform cross-modal translation from "in-the-wild'' monologue speech of a single speaker to their hand and arm motion. We train on unlabeled videos for which we only have noisy pseudo ground truth from an automatic pose detection system. Our proposed model significantly outperforms baseline methods in a quantitative comparison. To support research toward obtaining a computational understanding of the relationship between gesture and speech, we release a large video dataset of person-specific gestures. The project website with video, code and data can be found at http://people.eecs.berkeley.edu/~shiry/speech2gesture .

cs.CV↗

arXiv2019

PHT-bot: Deep-Learning based system for automatic risk stratification of COPD patients based upon signs of Pulmonary Hypertension

Chronic Obstructive Pulmonary Disease (COPD) is a leading cause of morbidity and mortality worldwide. Identifying those at highest risk of deterioration would allow more effective distribution of preventative and surveillance resources. Secondary pulmonary hypertension is a manifestation of advanced COPD, which can be reliably diagnosed by the main Pulmonary Artery (PA) to Ascending Aorta (Ao) ratio. In effect, a PA diameter to Ao diameter ratio of greater than 1 has been demonstrated to be a reliable marker of increased pulmonary arterial pressure. Although clinically valuable and readily visualized, the manual assessment of the PA and the Ao diameters is time consuming and under-reported. The present study describes a non invasive method to measure the diameters of both the Ao and the PA from contrast-enhanced chest Computed Tomography (CT). The solution applies deep learning techniques in order to select the correct axial slice to measure, and to segment both arteries. The system achieves test Pearson correlation coefficient scores of 93% for the Ao and 92% for the PA. To the best of our knowledge, it is the first such fully automated solution.

cs.CV↗

arXiv2017

Language Generation with Recurrent Generative Adversarial Networks without Pre-training

Generative Adversarial Networks (GANs) have shown great promise recently in image generation. Training GANs for language generation has proven to be more difficult, because of the non-differentiable nature of generating text with recurrent neural networks. Consequently, past work has either resorted to pre-training with maximum-likelihood or used convolutional networks for generation. In this work, we show that recurrent neural networks can be trained to generate text with GANs from scratch using curriculum learning, by slowly teaching the model to generate sequences of increasing and variable length. We empirically show that our approach vastly improves the quality of generated sequences compared to a convolutional baseline.

cs.CL↗

arXiv2017

Compression Fractures Detection on CT

The presence of a vertebral compression fracture is highly indicative of osteoporosis and represents the single most robust predictor for development of a second osteoporotic fracture in the spine or elsewhere. Less than one third of vertebral compression fractures are diagnosed clinically. We present an automated method for detecting spine compression fractures in Computed Tomography (CT) scans. The algorithm is composed of three processes. First, the spinal column is segmented and sagittal patches are extracted. The patches are then binary classified using a Convolutional Neural Network (CNN). Finally a Recurrent Neural Network (RNN) is utilized to predict whether a vertebral fracture is present in the series of patches.

cs.CV↗

arXiv2016

Exact extreme value statistics at mixed order transitions

We study extreme value statistics (EVS) for spatially extended models exhibiting mixed order phase transitions (MOT). These are phase transitions which exhibit features common to both first order (discontinuity of the order parameter) and second order (diverging correlation length) transitions. We consider here the truncated inverse distance squared Ising (TIDSI) model which is a prototypical model exhibiting MOT, and study analytically the extreme value statistics of the domain lengths. The lengths of the domains are identically distributed random variables except for the global constraint that their sum equals the total system size $L$. In addition, the number of such domains is also a fluctuating variable, and not fixed. In the paramagnetic phase, we show that the distribution of the largest domain length $l_{\max}$ converges, in the large $L$ limit, to a Gumbel distribution. However, at the critical point (for a certain range of parameters) and in the ferromagnetic phase, we show that the fluctuations of $l_{\max}$ are governed by novel distributions which we compute exactly. Our main analytical results are verified by numerical simulations.

cond-mat.stat-mech↗

arXiv2014

Mixed order transition and condensation in exactly soluble one dimensional spin model

Mixed order phase transitions (MOT), which display discontinuous order parameter and diverging correlation length, appear in several seemingly unrelated settings ranging from equilibrium models with long-range interactions to models far from thermal equilibrium. In a recent paper [1] an exactly soluble spin model with long-range interactions that exhibits MOT was introduced and analyzed both by a grand canonical calculation and a renormalization group analysis. The model was shown to lay a bridge between two classes of one dimensional models exhibiting MOT, namely between spin models with inverse distance square interactions and surface depinning models. In this paper we elaborate on the calculations done in [1]. We also analyze the model in the canonical ensemble, which yields a better insight into the mechanism of MOT. In addition, we generalize the model to include Potts and general Ising spins, and also consider a broader class of interactions which decay with distance with a power law different from 2.

cond-mat.stat-mech↗

arXiv2013

Mixed order phase transition in a one dimensional model

We introduce and analyze an exactly soluble one-dimensional Ising model with long range interactions which exhibits a mixed order transition (MOT), namely a phase transition in which the order parameter is discontinuous as in first order transitions while the correlation length diverges as in second order transitions. Such transitions are known to appear in a diverse classes of models which are seemingly unrelated. The model we present serves as a link between two classes of models which exhibit MOT in one dimension, namely, spin models with a coupling constant which decays as the inverse distance squared and models of depinning transitions, thus making a step towards a unifying framework.

cond-mat.stat-mech↗

arXiv2012

Denaturation of Circular DNA: Supercoils and Overtwist

The denaturation transition of circular DNA is studied within a Poland-Scheraga type approach, generalized to account for the fact that the total linking number (LK), which measures the number of windings of one strand around the other, is conserved. In the model the LK conservation is maintained by invoking both overtwisting and writhing (supercoiling) mechanisms. This generalizes previous studies which considered each mechanism separately. The phase diagram of the model is analyzed as a function of the temperature and the elastic constant $κ$ associated with the overtwisting energy for any given loop entropy exponent, $c$. As is the case where the two mechanisms apply separately, the model exhibits no denaturation transition for $c \le 2$. For $c>2$ and $κ=0$ we find that the model exhibits a first order transition. The transition becomes of higher order for any $κ>0$. We also calculate the contribution of the two mechanisms separately in maintaining the conservation of the linking number and find that it is weakly dependent on the loop exponent $c$.

cond-mat.stat-mech↗

arXiv2012

Macroscopic loop formation in circular DNA denaturation

The statistical mechanics of DNA denaturation under fixed linking number is qualitatively different from that of the unconstrained DNA. Quantitatively different melting scenarios are reached from two alternative assumptions, namely, that the denatured loops are formed in expense of 1) overtwist, 2) supercoils. Recent work has shown that the supercoiling mechanism results in a BEC-like picture where a macroscopic loop appears at Tc and grows steadily with temperature, while the nature of the denatured phase for the overtwisting case has not been studied. By extending an earlier result, we show here that a macroscopic loop appears in the overtwisting scenario as well. We calculate its size as a function of temperature and show that the fraction of the total sum of microscopic loops decreases above Tc, with a cusp at the critical point.

cond-mat.stat-mech↗

arXiv2011

Denaturation of Circular DNA: Supercoil Mechanism

The denaturation transition which takes place in circular DNA is analyzed by extending the Poland-Scheraga model to include the winding degrees of freedom. We consider the case of a homopolymer whereby the winding number of the double stranded helix, released by a loop denaturation, is absorbed by \emph{supercoils}. We find that as in the case of linear DNA, the order of the transition is determined by the loop exponent $c$. However the first order transition displayed by the PS model for $c>2$ in linear DNA is replaced by a continuous transition with arbitrarily high order as $c$ approaches 2, while the second-order transition found in the linear case in the regime $1<c\le2$ disappears. In addition, our analysis reveals that melting under fixed linking number is a \emph{condensation transition}, where the condensate is a macroscopic loop which appears above the critical temperature.

cond-mat.stat-mech↗

arXiv2008

Dynamics of DNA Melting

The dynamics of loops at the DNA denaturation transition is studied. A scaling argument is used to evaluate the asymptotic behavior of the autocorrelation function of the state of complementary bases (either open or closed). The long-time asymptotic behavior of the autocorrelation function is expressed in terms of the entropy exponent, c, of a loop. The validity of the scaling argument is tested using a microscopic model of an isolated loop and a toy model of interacting loops. This suggests a method for measuring the entropy exponent using single-molecule experiments such as florescence correlation spectroscopy.

cond-mat.stat-mech↗