arXiv subjects

Ikuro Sato

Publications and source records attributed to Ikuro Sato.

28 records · Page 2Linked to original sources

arXiv2019

Breaking Inter-Layer Co-Adaptation by Classifier Anonymization

This study addresses an issue of co-adaptation between a feature extractor and a classifier in a neural network. A naive joint optimization of a feature extractor and a classifier often brings situations in which an excessively complex feature distribution adapted to a very specific classifier degrades the test performance. We introduce a method called Feature-extractor Optimization through Classifier Anonymization (FOCA), which is designed to avoid an explicit co-adaptation between a feature extractor and a particular classifier by using many randomly-generated, weak classifiers during optimization. We put forth a mathematical proposition that states the FOCA features form a point-like distribution within the same class in a class-separable fashion under special conditions. Real-data experiments under more general conditions provide supportive evidences.

cs.LG↗

arXiv2018

Canonical and Compact Point Cloud Representation for Shape Classification

We present a novel compact point cloud representation that is inherently invariant to scale, coordinate change and point permutation. The key idea is to parametrize a distance field around an individual shape into a unique, canonical, and compact vector in an unsupervised manner. We firstly project a distance field to a $4$D canonical space using singular value decomposition. We then train a neural network for each instance to non-linearly embed its distance field into network parameters. We employ a bias-free Extreme Learning Machine (ELM) with ReLU activation units, which has scale-factor commutative property between layers. We demonstrate the descriptiveness of the instance-wise, shape-embedded network parameters by using them to classify shapes in $3$D datasets. Our learning-based representation requires minimal augmentation and simple neural networks, where previous approaches demand numerous representations to handle coordinate change and point permutation.

cs.CV↗

arXiv2017

Binary-decomposed DCNN for accelerating computation and compressing model without retraining

Recent trends show recognition accuracy increasing even more profoundly. Inference process of Deep Convolutional Neural Networks (DCNN) has a large number of parameters, requires a large amount of computation, and can be very slow. The large number of parameters also require large amounts of memory. This is resulting in increasingly long computation times and large model sizes. To implement mobile and other low performance devices incorporating DCNN, model sizes must be compressed and computation must be accelerated. To that end, this paper proposes Binary-decomposed DCNN, which resolves these issues without the need for retraining. Our method replaces real-valued inner-product computations with binary inner-product computations in existing network models to accelerate computation of inference and decrease model size without the need for retraining. Binary computations can be done at high speed using logical operators such as XOR and AND, together with bit counting. In tests using AlexNet with the ImageNet classification task, speed increased by a factor of 1.79, models were compressed by approximately 80%, and increase in error rate was limited to 1.20%. With VGG-16, speed increased by a factor of 2.07, model sizes decreased by 81%, and error increased by only 2.16%.

cs.CV↗

arXiv2015

APAC: Augmented PAttern Classification with Neural Networks

Deep neural networks have been exhibiting splendid accuracies in many of visual pattern classification problems. Many of the state-of-the-art methods employ a technique known as data augmentation at the training stage. This paper addresses an issue of decision rule for classifiers trained with augmented data. Our method is named as APAC: the Augmented PAttern Classification, which is a way of classification using the optimal decision rule for augmented data learning. Discussion of methods of data augmentation is not our primary focus. We show clear evidences that APAC gives far better generalization performance than the traditional way of class prediction in several experiments. Our convolutional neural network model with APAC achieved a state-of-the-art accuracy on the MNIST dataset among non-ensemble classifiers. Even our multilayer perceptron model beats some of the convolutional models with recently invented stochastic regularization techniques on the CIFAR-10 dataset.

cs.CV↗

arXiv2015

Pairwise Rotation Hashing for High-dimensional Features

Binary Hashing is widely used for effective approximate nearest neighbors search. Even though various binary hashing methods have been proposed, very few methods are feasible for extremely high-dimensional features often used in visual tasks today. We propose a novel highly sparse linear hashing method based on pairwise rotations. The encoding cost of the proposed algorithm is $\mathrm{O}(n \log n)$ for n-dimensional features, whereas that of the existing state-of-the-art method is typically $\mathrm{O}(n^2)$. The proposed method is also remarkably faster in the learning phase. Along with the efficiency, the retrieval accuracy is comparable to or slightly outperforming the state-of-the-art. Pairwise rotations used in our method are formulated from an analytical study of the trade-off relationship between quantization error and entropy of binary codes. Although these hashing criteria are widely used in previous researches, its analytical behavior is rarely studied. All building blocks of our algorithm are based on the analytical solution, and it thus provides a fairly simple and efficient procedure.

cs.CV↗

arXiv2013

The atomic-level mechanism underlying the functionality of aquaporin-0

So far, more than 82,000 protein structures have been reported in the Protein Data Bank, but the driving force and structures that allow for protein functions have not been elucidated at the atomic level for even one protein. We have been able to clarify that the inter-subunit hydrophobic interaction driving the electrostatic opening of the pore in aquaporin 0 (AQP0). Aquaporins are membrane channels for water and small non-ionic solutes found in animals, plants, and microbes. The structures of aquaporins have high homology and consist of homotetramers, each monomer of which has one pore for a water channel. Each pore has two narrow portions: one is the narrowest constriction region consisting of aromatic residues and an arginine (ar/R), and another is two asparagine-proline-alanine (NPA) homolog portions. Here we show that an inter-subunit hydrophobic interaction in AQP0 drives a stick portion consisting of four amino acids toward the pore and the tip of the stick portion, consisting of a nitrogen atom, opens the pore: that movement is the swing mechanism (this http URL). The energetics and conformational change of amino acids participating in the swing mechanism confirm this view. The swing mechanism in which inter-subunit hydrophobic interactions in the tetramer drive the on-off switching of the pore explains why aquaporins consist of tetramers. Here, we report that experimental and molecular dynamics findings using various mutants support this view of the swing mechanism. The finding that mutants of amino acids in AQP2 corresponding to the stick of the swing mechanism cause severe recessive nephrogenic diabetes insipidus (NDI) demonstrates the critical role of the swing mechanism for the aquaporin function. We report first that the inter-subunit hydrophobic interaction in aquaporin 0 drives the electrostatic opening of the aquaporin pore at the atomic level.

q-bio.BM↗

arXiv2007

Fitting two nucleons inside a box: exponentially suppressed corrections to the Luscher's formula

Scattering observables can be computed in lattice field theory by measuring the volume dependence of energy levels of two particle states. The dominant volume dependence, proportional to inverse powers of the volume, is determined by the phase shifts. This universal relation (\Lu's formula) between energy levels and phase shifts is distorted by corrections which, in the large volume limit, are exponentially suppressed. They may be sizable, however, for the volumes used in practice and they set a limit on how small the lattice can be in these studies. We estimate these corrections, mostly in the case of two nucleons. Qualitatively, we find that the exponentially suppressed corrections are proportional to the {\it square} of the potential (or to terms suppressed in the chiral expansion) and the effect due to pions going ``around the world'' vanishes. Quantitatively, the size of the lattice should be greater than $\approx(5 {fm})^3$ in order to keep finite volume corrections to the phase less than $1^\circ$ for realistic pion mass.

hep-lat↗

arXiv2006

Finite volume corrections to pi-pi scattering

Lattice QCD studies of hadron-hadron interactions are performed by computing the energy levels of the system in a finite box. The shifts in energy levels proportional to inverse powers of the volume are related to scattering parameters in a model independent way. In addition, there are non-universal exponentially suppressed corrections that distort this relation. These terms are proportional to exp(-m_pi L) and become relevant as the chiral limit is approached. In this paper we report on a one-loop chiral perturbation theory calculation of the leading exponential corrections in the case of I=2 pi-pi scattering near threshold.

hep-lat↗

arXiv2005

Combining Quark and Link Smearing to Improve Extended Baryon Operators

The effects of Gaussian quark-field smearing and analytic stout-link smearing on the correlations of gauge-invariant extended baryon operators are studied. Gaussian quark-field smearing substantially reduces contributions from the short wavelength modes of the theory, while stout-link smearing significantly reduces the noise from the stochastic evaluations. The use of gauge-link smearing is shown to be crucial for baryon operators constructed of covariantly-displaced quark fields. Preferred smearing parameters are determined for a lattice spacing a_s ~ 0.1 fm.

hep-lat↗

arXiv2005

Clebsch-Gordan Construction of Lattice Interpolating Fields for Excited Baryons

Large sets of baryon interpolating field operators are developed for use in lattice QCD studies of baryons with zero momentum. Operators are classified according to the double-valued irreducible representations of the octahedral group. At first, three-quark smeared, local operators are constructed for each isospin and strangeness and they are classified according to their symmetry with respect to exchange of Dirac indices. Nonlocal baryon operators are formulated in a second step as direct products of the spinor structures of smeared, local operators together with gauge-covariant lattice displacements of one or more of the smeared quark fields. Linear combinations of direct products of spinorial and spatial irreducible representations are then formed with appropriate Clebsch-Gordan coefficients of the octahedral group. The construction attempts to maintain maximal overlap with the continuum SU(2) group in order to provide a physically interpretable basis. Nonlocal operators provide direct couplings to states that have nonzero orbital angular momentum.

hep-lat↗