Welcome to Byte Size Arxiv

Deep meta-learning for the selection of accurate ultrasound based breast mass classifier

Therefore, in practice it would be beneficial to select the classification method based on the appearance of the particular US image. Our preliminary results demonstrate that meta-learning techniques can be used to improve the performance of the standard classifiers based on handcrafted features. With the proposed meta-learning based approach, we achieved the area under the receiver operating characteristic curve of 0.95 and accuracy of 0.91. Standard classification methods based on handcrafted morphological and texture features have achieved good performance in breast mass differentiation in ultrasound (US). In comparison to deep neural networks, commonly perceived as "black-box" models, classical techniques are based on features that have well-understood medical and physical interpretation. However, classifiers based on morphological features commonly underperform in the presence of the shadowing artifact and ill-defined mass borders, while texture based classifiers may fail when the US image is too noisy. Therefore, in practice it would be beneficial to select the classification method based on the appearance of the particular US image. In this work, we develop a deep meta-network that can automatically process input breast mass US images and recommend whether to apply the shape or texture based classifier for the breast mass differentiation. Our preliminary results demonstrate that meta-learning techniques can be used to improve the performance of the standard classifiers based on handcrafted features. With the proposed meta-learning based approach, we achieved the area under the receiver operating characteristic curve of 0.95 and accuracy of 0.91.

Authors: Michal Byra, Piotr Karwat, Ivan Ryzhankow, Piotr Komorowski, Ziemowit Klimonda, Lukasz Fura, Anna Pawlowska, Norbert Zolek, Jerzy Litniewski.

Protostellar collapse simulations in spherical geometry with dust coagulation and fragmentation

It solves the gas-dust hydrodynamics in a spherical geometry and the coagulation/fragmentation equation. It also computes the ionization state of the cloud and the Ohmic, ambipolar and Hall resistivities. At high density, we find that the coagulated distribution is unaffected by the initial choice of dust distribution. It is also found to be negligible for icy grains. When fragmentation occurs, it strongly affects the magnetic resistivities profiles. The onset and feedback of fragmentation remains uncertain and its modeling should be further investigated.

We model the coagulation and fragmentation of dust grains during the protostellar collapse with our newly developed shark code. It solves the gas-dust hydrodynamics in a spherical geometry and the coagulation/fragmentation equation. It also computes the ionization state of the cloud and the Ohmic, ambipolar and Hall resistivities. We find that the dust size distribution evolves significantly during the collapse, large grain formation being controlled by the turbulent differential velocity. When turbulence is included, only ambipolar diffusion remains efficient at removing the small grains from the distribution, brownian motion is only efficient as a standalone process. The macroscopic gas-dust drift is negligible for grain growth and only dynamically significant near the first Larson core. At high density, we find that the coagulated distribution is unaffected by the initial choice of dust distribution. Strong magnetic fields are found to enhance the small grains depletion, causing an important increase of the ambipolar diffusion. This hints that the magnetic field strength could be regulated by the small grain population during the protostellar collapse. Fragmentation could be effective for bare silicates, but its modeling relies on the choice of ill-constrained parameters. It is also found to be negligible for icy grains. When fragmentation occurs, it strongly affects the magnetic resistivities profiles. Dust coagulation is a critical process that needs to be fully taken into account during the protostellar collapse. The onset and feedback of fragmentation remains uncertain and its modeling should be further investigated.

Authors: Ugo Lebreuilly, Valentin Vallucci-Goy, Vincent Guillet, Maxime Lombart, Pierre Marchand.

Additive Combinatorics in Groups and Geometric Combinatorics on Spheres

We review some of the main known results in each area, mention several open questions, and discuss some connections among these four interesting topics. We embark on a tour that takes us through four closely related topics: the dual concepts of independence and spanning in finite abelian groups and the analogous dual concepts of designs and distance sets on spheres. We review some of the main known results in each area, mention several open questions, and discuss some connections among these four interesting topics.

Authors: Bela Bajnok.

When to Laugh and How Hard? A Multimodal Approach to Detecting Humor and its Intensity

Our model is capable of recognizing whether an utterance is humorous or not and assess the intensity of it. We evaluate the model on episodes the model has not been exposed to during the training phase.

Prerecorded laughter accompanying dialog in comedy TV shows encourages the audience to laugh by clearly marking humorous moments in the show. We present an approach for automatically detecting humor in the Friends TV show using multimodal data. Our model is capable of recognizing whether an utterance is humorous or not and assess the intensity of it. We use the prerecorded laughter in the show as annotation as it marks humor and the length of the audience's laughter tells us how funny a given joke is. We evaluate the model on episodes the model has not been exposed to during the training phase. Our results show that the model is capable of correctly detecting whether an utterance is humorous 78% of the time and how long the audience's laughter reaction should last with a mean absolute error of 600 milliseconds.

Authors: Khalid Alnajjar, Mika Hämäläinen, Jörg Tiedemann, Jorma Laaksonen, Mikko Kurimo.

Accurate waveform models for gravitational-wave astrophysics: synergetic approaches from analytical relativity

Gravitational-wave (GW) astrophysics is a field in full blossom. [Abridged.] Gravitational-wave (GW) astrophysics is a field in full blossom. Since the landmark detection of GWs from a binary black hole on September 14th 2015, several compact-object binaries have been reported by the LIGO-Virgo collaboration. Such events carry astrophysical and cosmological information ranging from an understanding of how black holes and neutron stars are formed, what neutron stars are composed of, how the Universe expands, and allow testing general relativity in the highly-dynamical strong-field regime. It is the goal of GW astrophysics to extract such information as accurately as possible. Yet, this is only possible if the tools and technology used to detect and analyze GWs are advanced enough. A key aspect of GW searches are waveform models, which encapsulate our best predictions for the gravitational radiation under a certain set of parameters, and that need to be cross-correlated with data to extract GW signals. Waveforms must be very accurate to avoid missing important physics in the data, which might be the key to answer the fundamental questions of GW astrophysics. The continuous improvements of the current LIGO-Virgo detectors, the development of next-generation ground-based detectors such as the Einstein Telescope or the Cosmic Explorer, as well as the development of the Laser Interferometer Space Antenna (LISA), demand accurate waveform models. [Abridged.]

Authors: Andrea Antonelli.

Thermalization of Classical Weakly Nonintegrable Many-Body Systems

We devote our studies to the subject of weakly nonintegrable dynamics of systems with a macroscopic number of degrees of freedom. We solve these challenges by performing numerical tests using computationally efficient model - unitary maps. The great advantage of unitary maps for numerical applications is time-discrete error-free evolution. To demonstrate the scope of obtained results we report on the application of the developed framework to Hamiltonian systems.

We devote our studies to the subject of weakly nonintegrable dynamics of systems with a macroscopic number of degrees of freedom. Our main points of interest are the relations between the timescales of thermalization and the timescales of chaotization; the choice of appropriate observables and the structure of equations coupling them; identifying the classes of weakly nonintegrable dynamics and developing tools to diagnose the properties specific to such classes. We discuss the traditional in the field methods for thermalization timescale computation and employ them to study the scaling the timescale with the proximity to the integrable limit. We then elaborate on a novel framework based on the full Lyapunov spectra computation for large systems as a powerful tool for the characterization of weak nonintegrability. In particular, the Lyapunov spectrum scaling offers a quantitative description allowing us to infer the structure of the underlying network of observables. Proximity to integrable limit is associated with the rapid growth of thermalization timescales and, thus, potential numerical challenges. We solve these challenges by performing numerical tests using computationally efficient model - unitary maps. The great advantage of unitary maps for numerical applications is time-discrete error-free evolution. We use these advantages to perform large timescale and system size computations in extreme proximity to the integrable limit. To demonstrate the scope of obtained results we report on the application of the developed framework to Hamiltonian systems.

Authors: Merab Malishava.

Analysing the effectiveness of a generative model for semi-supervised medical image segmentation

Image segmentation is important in medical imaging, providing valuable, quantitative information for clinical decision-making in diagnosis, therapy, and intervention. The state-of-the-art in automated segmentation remains supervised learning, employing discriminative models such as U-Net. Recently, generative models have been proposed for semantic segmentation, as they make an attractive choice for SSL. Their ability to capture the joint distribution over input images and output label maps provides a natural way to incorporate information from unlabelled images. Image segmentation is important in medical imaging, providing valuable, quantitative information for clinical decision-making in diagnosis, therapy, and intervention. The state-of-the-art in automated segmentation remains supervised learning, employing discriminative models such as U-Net. However, training these models requires access to large amounts of manually labelled data which is often difficult to obtain in real medical applications. In such settings, semi-supervised learning (SSL) attempts to leverage the abundance of unlabelled data to obtain more robust and reliable models. Recently, generative models have been proposed for semantic segmentation, as they make an attractive choice for SSL. Their ability to capture the joint distribution over input images and output label maps provides a natural way to incorporate information from unlabelled images. This paper analyses whether deep generative models such as the SemanticGAN are truly viable alternatives to tackle challenging medical image segmentation problems. To that end, we thoroughly evaluate the segmentation performance, robustness, and potential subgroup disparities of discriminative and generative segmentation methods when applied to large-scale, publicly available chest X-ray datasets.

Authors: Margherita Rosnati, Fabio De Sousa Ribeiro, Miguel Monteiro, Daniel Coelho de Castro, Ben Glocker.

Using U-Net Network for Efficient Brain Tumor Segmentation in MRI Images

Magnetic Resonance Imaging (MRI) is the most commonly used non-intrusive technique for medical image acquisition. Brain tumor segmentation is the process of algorithmically identifying tumors in brain MRI scans. Moreover, no additional data augmentation step is required.

Magnetic Resonance Imaging (MRI) is the most commonly used non-intrusive technique for medical image acquisition. Brain tumor segmentation is the process of algorithmically identifying tumors in brain MRI scans. While many approaches have been proposed in the literature for brain tumor segmentation, this paper proposes a lightweight implementation of U-Net. Apart from providing real-time segmentation of MRI scans, the proposed architecture does not need large amount of data to train the proposed lightweight U-Net. Moreover, no additional data augmentation step is required. The lightweight U-Net shows very promising results on BITE dataset and it achieves a mean intersection-over-union (IoU) of 89% while outperforming the standard benchmark algorithms. Additionally, this work demonstrates an effective use of the three perspective planes, instead of the original three-dimensional volumetric images, for simplified brain tumor segmentation.

Authors: Jason Walsh, Alice Othmani, Mayank Jain, Soumyabrata Dev.

Orbital liquid in the $e_g$ orbital Hubbard model in $d=\infty$ dimensions

While the local Coulomb interaction $U$ is invariant for each basis of orthogonal orbitals, the form of the kinetic energy depends on the orbital basis and takes the most symmetric form for the so-called complex-orbital basis. Characteristically, with respect to this basis, the model has two hopping channels, one that is orbital-flavor conserving, and a second one that is orbital-flavor non-conserving. We show that the noninteracting electronic structure consists of two nondegenerate bands of plane-wave real-orbital single-particle states for which the orbital depends on the wave vector. The \textit{orbital liquid} state is obtained by filling these two bands up to the same Fermi energy. The latter feature is shown to be specific for $d=\infty$, being of mathematical nature due to the exponential tails in the density of states. We demonstrate that the three-dimensional $e_g$ orbital Hubbard model can be generalized to arbitrary dimension $d$, and that the form of the result is determined uniquely by the requirements that (i) the two-fold degeneracy of the $e_g$ orbital be retained, and (ii) the cubic lattice be turned into a hypercubic lattice. While the local Coulomb interaction $U$ is invariant for each basis of orthogonal orbitals, the form of the kinetic energy depends on the orbital basis and takes the most symmetric form for the so-called complex-orbital basis. Characteristically, with respect to this basis, the model has two hopping channels, one that is orbital-flavor conserving, and a second one that is orbital-flavor non-conserving. We show that the noninteracting electronic structure consists of two nondegenerate bands of plane-wave real-orbital single-particle states for which the orbital depends on the wave vector. Due to the latter feature each band is unpolarized at any filling, and has a non-Gaussian density of states at $d=\infty$. The \textit{orbital liquid} state is obtained by filling these two bands up to the same Fermi energy. We investigate the $e_g$ orbital Hubbard model in the limit $d\to\infty$, treating the on-site Coulomb interaction $U$ within the Gutzwiller approximation, thus determining the correlation energy of the orbital liquid and the (disordered) para-orbital states. (...) We show that the orbital liquid is the ground state everywhere in the $(n,U)$ phase diagram except close to half-filling at sufficiently large $U$, where ferro-orbital order with real orbitals occupied is favored. The latter feature is shown to be specific for $d=\infty$, being of mathematical nature due to the exponential tails in the density of states.

Authors: Louis Felix Feiner, Andrzej M. Oleś.

Faster Adaptive Momentum-Based Federated Methods for Distributed Composition Optimization

Composition optimization recently appears in many machine learning applications such as meta learning and reinforcement learning. Moreover, we provide a solid theoretical analysis for our algorithms under non-i.i.d. Specifically, our algorithms obtain lower sample complexity of $\tilde{O}(\epsilon^{-3})$ with lower communication complexity of $\tilde{O}(\epsilon^{-2})$ in finding an $\epsilon$-stationary point. We conduct the experiments on robust federated learning and distributed meta learning tasks to demonstrate efficiency of our algorithms.

Composition optimization recently appears in many machine learning applications such as meta learning and reinforcement learning. Recently many composition optimization algorithms have been proposed and studied, however, few adaptive algorithm considers the composition optimization under the distributed setting. Meanwhile, the existing distributed composition optimization methods still suffer from high sample and communication complexities. In the paper, thus, we develop a class of faster momentum-based federated compositional gradient descent algorithms (i.e., MFCGD and AdaMFCGD) to solve the nonconvex distributed composition problems, which builds on the momentum-based variance reduced and local-SGD techniques. In particular, our adaptive algorithm (i.e., AdaMFCGD) uses a unified adaptive matrix to flexibly incorporate various adaptive learning rates. Moreover, we provide a solid theoretical analysis for our algorithms under non-i.i.d. setting, and prove our algorithms obtain a lower sample and communication complexities simultaneously than the existing federated compositional algorithms. Specifically, our algorithms obtain lower sample complexity of $\tilde{O}(\epsilon^{-3})$ with lower communication complexity of $\tilde{O}(\epsilon^{-2})$ in finding an $\epsilon$-stationary point. We conduct the experiments on robust federated learning and distributed meta learning tasks to demonstrate efficiency of our algorithms.

Authors: Feihu Huang.