Skip to main content
main-content

Über dieses Buch

The four volume set LNCS 9947, LNCS 9948, LNCS 9949, and LNCS 9950 constitues the proceedings of the 23rd International Conference on Neural Information Processing, ICONIP 2016, held in Kyoto, Japan, in October 2016. The 296 full papers presented were carefully reviewed and selected from 431 submissions. The 4 volumes are organized in topical sections on deep and reinforcement learning; big data analysis; neural data analysis; robotics and control; bio-inspired/energy efficient information processing; whole brain architecture; neurodynamics; bioinformatics; biomedical engineering; data mining and cybersecurity workshop; machine learning; neuromorphic hardware; sensory perception; pattern recognition; social networks; brain-machine interface; computer vision; time series analysis; data-driven approach for extracting latent features; topological and graph based clustering methods; computational intelligence; data mining; deep neural networks; computational and cognitive neurosciences; theory and algorithms.

Inhaltsverzeichnis

Frontmatter

Applications

Frontmatter

Classifying Human Activities with Temporal Extension of Random Forest

Sensor-Based Human Activity Recognition (HAR) is a study of recognizing the human’s activities by using the data captured from wearable sensors. Avail the temporal information from the sensors, a modified version of random forest is proposed to preserve the temporal information, and harness them in classifying the human activities. The proposed algorithm is tested on 7 public HAR datasets. Promising results are reported, with an average classification accuracy of ~98 %.

Shih Yin Ooi, Shing Chiang Tan, Wooi Ping Cheah

Echo State Network Ensemble for Human Motion Data Temporal Phasing: A Case Study on Tennis Forehands

Temporal phasing analysis is integral to ubiquitous/“smart” coaching devices and sport science. This study presents a novel approach to autonomous temporal phasing of human motion from captured tennis activity (3D data, 66 time-series). Compared to the optimised Echo State Network (ESN) model achieving 85 % classification accuracy, the ESN ensemble system demonstrates improved classification of 95 % and 100 % accurate phasing state transitions for previously unseen motions without requiring ball impact information. The ESN ensemble model is robust to low-sampling rates (50 Hz) and unbalanced data sets containing incomplete data time-series. The demonstrated achievements are applicable to exergames, augmented coaching and rehabilitation systems advancements by enabling automated qualitative analysis of motion data and generating feedback to aid motor skill and technique improvements.

Boris Bačić

Unregistered Bosniak Classification with Multi-phase Convolutional Neural Networks

Deep learning has been a growing trend in various fields of natural image classification as it performs state-of-the-art result on several challenging tasks. Despite its success, deep learning applied to medical image analysis has not been wholly explored. In this paper, we study on convolutional neural network (CNN) architectures applied to a Bosniak classification problem to classify Computed Tomography images into five Bosniak classes. We use a new medical image dataset called as the Bosniak classification dataset which will be fully introduced in this paper. For this data set, we employ a multi-phase CNN approach to predict classification accuracy. We also discuss the representation power of CNN compared to previously developed features (Garbor features) in medical image. In our experiment, we use data combination method to enlarge the data set to avoid overfitting problem in multi-phase medical imaging system. Using multi-phase CNN and data combination method we proposed, we have achieved 48.9 % accuracy on our test set, which improves the hand-crafted features by 11.9 %.

Myunggi Lee, Hyeogjin Lee, Jiyong Oh, Hak Jong Lee, Seung Hyup Kim, Nojun Kwak

Direct Estimation of Wrist Joint Angular Velocities from Surface EMGs by Using an SDNN Function Approximator

The present paper proposes a method for estimating joint angular velocities from multi-channel surface electromyogram (sEMG) signals. This method uses a selective desensitization neural network (SDNN) as a function approximator that learns the relation between integrated sEMG signals and instantaneous joint angular velocities. A comparison experiment with a Kalman filter model shows that this method can estimate wrist angular velocities in real time with high accuracy, especially during rapid motion.

Kazumasa Horie, Atsuo Suemitsu, Tomohiro Tanno, Masahiko Morita

Data Analysis of Correlation Between Project Popularity and Code Change Frequency

Github is a source code management platform with social networking features that help increase the popularity of a project. The features of the GitHub like watch, star, fork and pull requests help make a project popular among the developers, in addition to enabling them to work on the code together. In this work, we study the relation between the project popularity and the continual code changes made to a GitHub project. The correlation is found by using the metrics such as the number of watchers, pull requests, and the number of commits. We correlate the time series of code change frequency with the time series of project popularity. As a result, we have found that projects with at least 1500 watchers each month have a strong positive correlation between the project popularity and frequency of code changes. We have also found that the number of pull requests is 73.2 % more important to the popularity of a project than the number of watchers.

Dabeeruddin Syed, Jadran Sessa, Andreas Henschel, Davor Svetinovic

Hidden Space Neighbourhood Component Analysis for Cancer Classification

Neighbourhood component analysis (NCA) is a method for learning a distance metric which can maximize the classification performance of the K nearest neighbour (KNN) classifier. However, NCA suffers from the small size sample problem that the number of samples is much less than the number of features. To remedy this, this paper proposes a hidden space neighbourhood components analysis (HSNCA), which is a nonlinear extension of NCA. HSNCA first maps the data in the original space into a feature space by a set of nonlinear mapping functions, and then performs NCA in the feature space. Notably, the number of samples is equal to the number of features in the feature space. Thus, HSNCA can avoid the small size sample problem. Experimental results on DNA array datasets show that HSNCA is feasibility and efficiency.

Li Zhang, Xiaojuan Huang, Bangjun Wang, Fanzhang Li, Zhao Zhang

Prediction of Bank Telemarketing with Co-training of Mixture-of-Experts and MLP

Utilization of financial data becomes one of the important issues for user adaptive marketing on the bank service. The marketing is conducted based on personal information with various facts that affect a success (clients agree to accept financial instrument). Personal information can be collected continuously anytime if clients want to agree to use own information in case of opening an account in bank, but labeling all the data needs to pay a high cost. In this paper, focusing on this characteristics of financial data, we present a global-local co-training (GLCT) algorithm to utilize labeled and unlabeled data to construct better prediction model. We performed experiments using real-world data from Portuguese bank. Experiments show that GLCT performs well regardless of the ratio of initial labeled data. Through the series of iterating experiments, we obtained better results on various aspects.

Jae-Min Yu, Sung-Bae Cho

Prioritising Security Tests on Large-Scale and Distributed Software Development Projects by Using Self-organised Maps

Large-scale and distributed software development initiatives demand a systematic testing process in order to prevent failures. Significant amount of resources are usually allocated on testing. Like any development and designing task, testing activities have to be prioritised in order to efficiently validate the produced code. By using source code complexity measurement, Computational Intelligence and Image Processing techniques, this research presents a new approach to prioritise testing efforts on large-scale and distributed software projects. The proposed technique was validated by automatically highlighting sensitive code within the Linux device drivers source code base. Our algorithm was able to classify 3, 077 from 35, 091 procedures as critical code to be tested. We argue that the approach is general enough to prioritise test tasks of most critical large-scale and distributed developed software such as: Operating Systems, Enterprise Resource Planning and Content Management systems.

Marcos Alvares, Fernando Buarque de Lima Neto, Tshilidzi Marwala

Android Malware Detection Method Based on Function Call Graphs

With the rapid development of mobile Internet, mobile devices have been widely used in people’s daily life, which has made mobile platforms a prime target for malware attack. In this paper we study on Android malware detection method. We propose the method how to extract the structural features of android application from its function call graph, and then use the structure features to build classifier to classify malware. The experiment results show that structural features can effectively improve the performance of malware detection methods.

Yuxin Ding, Siyi Zhu, Xiaoling Xia

Proposal of Singular-Unit Restoration by Focusing on the Spatial Continuity of Topographical Statistics in Spectral Domain

An interferogram which interferometric synthetic aperture radar (InSAR) acquires includes singular points (SPs), which cause an unwrapping error. It is very important to remove the SP. We propose a filtering technique in order to eliminate the distortion around a SP. In this proposed filter, a complex-valued neural network (CVNN) learns the continuous changes of topographical statistics in the spectral domain. CVNN predicts the spectrum around a singular unit (SU), i.e., the four pixels constituting a SP, to restore the SU. The proposed method is so effective in the removal of the distortion at the SU that it allows us to generate a highly accurate digital elevation model (DEM).

Kazuhide Ichikawa, Akira Hirose

Inferring Users’ Gender from Interests: A Tag Embedding Approach

This paper studies the problem of gender prediction of users in social media using their interest tags. The challenge is that the tag feature vector is extremely sparse and short, i.e., less than 10 tags for each user. We present a novel conceptual class based method which enriches and centralizes the feature space. We first identify the discriminating tags based on the tag distribution. We then build the initial conceptual class by taking the advantage of the generalization and specification operations on these tags. For example, “Kobe” is a specialized instance of “basketball”. Finally, we model class expansion as a problem of computing the similarity between one tag and a set of tags in one conceptual class in the embedding space.We conduct extensive experiments on a real dataset from Sina Weibo. Results demonstrate that our proposed method significantly enhances the quality of the feature space and improves the performance of gender classification. Its accuracy reaches 82.25 % while that for the original tag vector is only 62.75 %.

Peisong Zhu, Tieyun Qian, Ming Zhong, Xuhui Li

Fast Color Quantization via Fuzzy Clustering

This comparative study employs several modified versions of the fuzzy c-means algorithm in image color reduction, with the aim of assessing their accuracy and efficiency. To assure equal chances for all algorithms, a common framework was established that preprocesses input images in terms of a preliminary color quantization, extraction of histogram and selection of frequently occurring colors of the image. Selected colors were fed to clustering by studied c-means algorithm variants. Besides the conventional fuzzy c-means (FCM) algorithm, the so-called generalized improved partition FCM algorithm, and several versions of the generalized suppressed FCM were considered. Accuracy was assessed by the average color difference between input and output images, while efficiency tests monitored the total runtime. All modified algorithms were found more accurate, and some suppressed models also faster than FCM.

László Szilágyi, Gellért Dénesi, Călin Enăchescu

Extended Dependency-Based Word Embeddings for Aspect Extraction

Extracting aspects from opinion reviews is an essential task of fine-grained sentiment analysis. In this paper, we introduce outer product of dependency-based word vectors and specialized features as representation of words. With such extended embeddings composed in recurrent neural networks, we make use of advantages of both word embeddings and traditional features. Evaluated on SemEval 2014 task 4 dataset, the proposed method outperform existing recurrent models based methods, achieving a result comparable with the state-of-the-art method. It shows that it is an effective way to achieve better extraction performance by improving word representations.

Xin Wang, Yuanchao Liu, Chengjie Sun, Ming Liu, Xiaolong Wang

Topological Order Discovery via Deep Knowledge Tracing

The goal of discovering topological order of skills is to generate a sequence of skills satisfying all prerequisite requirements. Very few previous studies have examined this task from knowledge tracing perspective. In this paper, we introduce a new task of discovering topological order of skills using students’ exercise performance and explore the utility of Deep Knowledge Tracing (DKT) to solve this task. The learned topological results can be used to improve students’ learning efficiency by providing students with personalized learning paths and predicting students’ future exercise performance. Experimental results demonstrate that our method is effective to generate reasonable topological order of skills.

Jiani Zhang, Irwin King

PTR: Phrase-Based Topical Ranking for Automatic Keyphrase Extraction in Scientific Publications

Automatic keyphrase extraction plays an important role for many information retrieval (IR) and natural language processing (NLP) tasks. Motivated by the facts that phrases have more semantic information than single words and a document consists of multiple semantic topics, we present PTR, a phrase-based topical ranking method for keyphrase extraction in scientific publications. Candidate keyphrases are divided into different topics by LDA and used as vertices in a phrase-based graph of the topic. We then decompose PageRank into multiple weighted-PageRank to rank phrases for each topic. Keyphrases are finally generated by selecting candidates according to their overall scores on all related topics. Experimental results show that PTR has good performance on several datasets.

Minmei Wang, Bo Zhao, Yihua Huang

Neural Network Based Association Rule Mining from Uncertain Data

In data mining, the U-Apriori algorithm is typically used for Association Rule Mining (ARM) from uncertain data. However, it takes too much time in finding frequent itemsets from large datasets. This paper proposes a novel algorithm based on Self-Organizing Map (SOM) clustering for ARM from uncertain data. It supports the feasibility of neural network for generating frequent itemsets and association rules effectively. We take transactions in which itemsets are associated with probabilities of occurrence. Each transaction is converted to an input vector under a probabilistic framework. SOM is employed to train these input vectors and visualize the relationship between the items in a database. Distance map based on the weights of winning neurons and support count of items is used as a criteria to prune data space. As shown in our experiments, the proposed SOM is a promising alternative to typical mining algorithms for ARM from uncertain data.

Sameen Mansha, Zaheer Babar, Faisal Kamiran, Asim Karim

Analysis and Knowledge Discovery by Means of Self-Organizing Maps for Gaia Data Releases

A billion stars: this is the approximate amount of visible objects estimated to be observed by the Gaia satellite, representing roughly 1 % of the objects in the Galaxy. It constitutes the biggest amount of data gathered to date: by the end of the mission, the data archive will exceed 1 Petabyte. Now, in order to process this data, the Gaia mission conceived the Data Processing and Analysis Consortium, which will apply data mining techniques such as Self-Organizing Maps. This paper shows a useful technique for source clustering, focusing on the development of an advanced visualization tool based on this technique.

Marco Antonio Álvarez, Carlos Dafonte, Daniel Garabato, Minia Manteiga

Computational and Cognitive Neurosciences

Frontmatter

The Impact of Adaptive Regularization of the Demand Predictor on a Multistage Supply Chain Simulation

The supply chain is difficult to control, which is representative of the bullwhip effect. Its behavior under the influence of the bullwhip effect is complex, and the cost and risk are increased. This study provides an application of online learning that is effective in large-scale data processing in a supply chain simulation. Because quality of solutions and agility are required in the management of the supply chain, we have adopted adaptive regularization learning. This is excellent from the viewpoint of speed and generalization of convergence and can be expected to stabilize supply chain behavior. In addition, because it is an online learning algorithm for evaluation of the bullwhip effect by computer simulation, it is easily applied to large-scale data from the viewpoint of the amount of calculation and memory size. The effectiveness of our approach was confirmed.

Fumiaki Saitoh

The Effect of Reward Information on Perceptual Decision-Making

Decision making can be treated as a two-step process involving sensory information and valuation of various options. However, the integration of value and sensory information at a neural level is still unclear. We used electroencephalography (EEG) to investigate the effect of reward information on perceptual decision making using two- alternative discriminating task. The reward information was signalled before the appearance of the stimuli. Our findings suggest that economic value acts as a top-down influence early in the decision epoch possibly shifting the evaluation criteria to a more favourable outcome.

Devu Mahesan, Manisha Chawla, Krishna P. Miyapuram

Doubting What to Eat: A Computational Model for Food Choice Using Different Valuing Perspectives

In this paper a computational model for the decision making process of food choices is presented that takes into account a number of aspects on which a decision can be based, for example, a temptation triggered by the food itself, a desire for food triggered by being hungry, valuing by the expected basic satisfaction feeling, and valuing by the expected goal satisfaction feeling.

Altaf H. Abro, Jan Treur

A Novel Graph Regularized Sparse Linear Discriminant Analysis Model for EEG Emotion Recognition

In this paper, a novel regression model, called graph regularized sparse linear discriminant analysis (GraphSLDA), is proposed to deal with EEG emotion recognition problem. GraphSLDA extends the conventional linear discriminant analysis (LDA) method by imposing a graph regularization and a sparse regularization on the transform matrix of LDA, such that it is able to simultaneously cope with sparse transform matrix learning while preserve the intrinsic manifold of the data samples. To cope with the EEG emotion recognition, we extract a set of frequency based EEG features to training the GraphSLDA model and also use it as EEG emotion classifier for testing EEG signals, in which we divide the raw EEG signals into five frequency bands, i.e., $$\delta $$δ, $$\theta $$θ, $$\alpha $$α, $$\beta $$β, and $$\gamma $$γ. To evaluate the proposed GraphSLDA model, we conduct experiments on the SEED database. The experimental results show that the proposed algorithm GraphSLDA is superior to the classic baselines.

Yang Li, Wenming Zheng, Zhen Cui, Xiaoyan Zhou

Information Maximization in a Feedforward Network Replicates the Stimulus Preference of the Medial Geniculate and the Auditory Cortex

Central auditory neurons exhibit a preference for complex features, such as frequency modulation and pitch. This study shows that the stimulus preference for these features can be replicated by a network model trained to maximize information transmission from input to output. The network contains three layers: input, first-output, and second-output. The first-output-layer neurons exhibit auditory-nerve neuron-like preferences, and the second-output-layer neurons exhibit a stimulus preference similar to that of cochlear nucleus, medial geniculate, and auditory cortical neurons. The features detected by the second-output-layer neurons reflect the statistical properties of the sounds used as input.

Takuma Tanaka

A Simple Visual Model Accounts for Drift Illusion and Reveals Illusory Patterns

Computational models of vision should not only be able to reproduce experimentally obtained results; such models should also be able to predict the input–output properties of vision. We assess whether a simple computational model of neurons in the Middle Temporal (MT) visual area proposed by the authors can account for illusory perception of “rotating drift patterns,” by which humans perceive illusory rotation (clockwise or counterclockwise) depending on the background luminance. Moreover, to predict whether a pattern causes visual illusion or not, we generate an enormous set of possible visual patterns as inputs to the MT model: $$ 8^{8} = 16,777,216, $$88=16,777,216, possible input patterns. Numerical quantities of model outputs by computer simulation for 88 inputs were used to estimate human illusory perception. Using psychophysical experiments, we show that the model prediction is consistent with human perception.

Daiki Nakamura, Shunji Satoh

An Internal Model of the Human Hand Affects Recognition of Graspable Tools

In this study, we validated a plausibility of a hypothesis that in the human brain an internal simulation of grasping contributes to tool recognition. Such an internal simulation must be performed by utilizing internal models of the human hand. An internal model corresponding to a geometrically transformed hand shape was retrained by an experimental paradigm we built. The retrained internal model of the dominant hand affected cognitive judgments of object size of tools used by the dominant hand and however did not influence these of tools used by the non-dominant hand. While, those results in the training condition of the non-dominant hand showed the reverse tendency of the former results. The above results indicate the plausibility of the hypothesis.

Masazumi Katayama, Yusuke Akimaru

Perceptual Representation of Material Quality: Adaptation to BRDF-Morphing Images

Perception of the material quality of a surface depends on its reflectance properties. Recent physiological studies reported the neural selectivity to glossy surfaces in the Inferior Temporal cortical areas [e.g., 1]. In the present study, we examine the hypothesis that basis neurons are selective to typical materials, and that the combinations of their responses are representative of a variety of natural materials. To assess the hypotheses, we performed a psychological experiment based on adaptation. If adaptation to a specific material is observed, the presence of neurons that are selective to the specific material is predicted. We performed adaptation tests with six typical material qualities including gloss, matte, metal and wood. We observed the adaptation to certain materials but not to some other materials. This result indicates the presence of basis neurons that are selective to materials, which is fundamentally important for understanding cortical representation of surface materials.

K. Kudou, K. Sakai

GPU-Accelerated Simulations of an Electric Stimulus and Neural Activities in Electrolocation

To understand mechanism of information processing by a neural network, it is important to well know a sensory stimulus. However, it is hard to examine details of a real stimulus received by an animal. Furthermore, it is too hard to simultaneously measure a received stimulus and neural activities of a neural system. We have studied the electrosensory system of an electric fish in electrolocation. It is also difficult to measure the electric stimulus received by an electric fish in the real environment and neural activities evoked by the electric stimulus. To address this issue, we have applied computational simulation. We developed the simulation software accelerated by a GPU to calculate various electric stimuli and neural activities of the electrosensory system using a GPU. This paper describes comparison of computation time between CPUs and a GPU in calculation of the electric field and the neural activities.

Kazuhisa Fujita, Yoshiki Kashimori

Analysis of Similarity and Differences in Brain Activities Between Perception and Production of Facial Expressions Using EEG Data and the NeuCube Spiking Neural Network Architecture

This paper is a feasibility study of using the NeuCube spiking neural network (SNN) architecture for modeling EEG brain data related to perceiving versus mimicking facial expressions. It is demonstrated that the proposed model can be used to study the similarity and differences between corresponding brain activities as complex spatio-temporal patterns. Two SNN models are created for each of the 7 basic emotions for a group of Japanese subjects, one when subjects are perceiving an emotional face and another, when the same subjects are mimicking this emotion. The evolved connectivity in the two models are then subtracted to study the differences. Analysis of the models trained on the collected EEG data shows greatest similarity in sadness, and least similarity in happiness and fear, where differences in the T6 EEG channel area were observed. The study, being based on the well-known mirror neuron concept in the brain, is the first to analyze and visualize similarity and differences as evolved spatio-temporal patterns in a brain-like SNN model.

Hideaki Kawano, Akinori Seo, Zohreh Gholami Doborjeh, Nikola Kasabov, Maryam Gholami Doborjeh

Self and Non-self Discrimination Mechanism Based on Predictive Learning with Estimation of Uncertainty

In this paper, we propose a model that can explain the mechanism of self and non-self discrimination. Infants gradually develop their abilities for self–other cognition through interaction with the environment. Predictive learning has been widely used to explain the mechanism of infants’ development. We hypothesized that infants’ cognitive abilities are developed through predictive learning and the uncertainty estimation of their sensory-motor inputs. We chose a stochastic continuous time recurrent neural network, which is a dynamical neural network model, to predict uncertainties as variances. From the perspective of cognitive developmental robotics, a predictive learning experiment with a robot was performed. The results indicate that training made the robot predict the regions related to its body more easily. We confirmed that self and non-self cognitive abilities might be acquired through predictive learning with uncertainty estimation.

Ryoichi Nakajo, Maasa Takahashi, Shingo Murata, Hiroaki Arie, Tetsuya Ogata

A Framework for Ontology Based Management of Neural Network as a Service

Neural networks proved extremely feasible for problems which are hard to solve by conventional computational algorithms due to excessive computational demand, as NP-hard problems, or even lack of a deterministic solution approach. In this paper we present a management framework for neural network objects based on ontology knowledge for the cloud-based neural network simulator N2Sky, which delivers neural network resources as a service on a world-wide basis. Core of this framework is the Neural Network Query Engine, N2Query, which allows users to specify their problem statements in form of natural language queries. It delivers a list of ranked N2Sky resources in return, providing solutions to these problems. The search algorithm applies a mapping process between a domain specific problem ontology and solution ontology.

Erich Schikuta, Abdelkader Magdy, A. Baith Mohamed

Computational Model of the Cerebellum and the Basal Ganglia for Interval Timing Learning

In temporal information processing, both the cerebellum and the basal ganglia play essential roles. In particular, for interval timing learning, the cerebellum exhibits temporally localized activity around the onset of the unconditioned stimulus, whereas the basal ganglia represents the passage of time by their ramping-up activity from the onset of the conditioned stimulus to that of the unconditioned stimulus. We present a unified computational model of the cerebellum and the basal ganglia for the interval timing learning task. We report that our model reproduces the localized activity in the cerebellum and the gradual increase of the activity in the basal ganglia. These results suggest that the cerebellum and the basal ganglia play different roles in temporal information processing.

Ohki Katakura, Tadashi Yamazaki

Bihemispheric Cerebellar Spiking Network Model to Simulate Acute VOR Motor Learning

The vestibuloocular reflex (VOR) is an adaptive control system. The cerebellar flocculus is intimately involved in the VOR adaptive motor control. The cerebellar flocculus has bihemispheric architecture and the several lines of unilateral lesion study indicated that each cerebellar hemisphere plays different roles in the leftward and rightward eye movement control and learning. However, roles of bihemispheric cerebellar architecture underlying the VOR motor learning have not been fully understood. Here we configure an anatomically/physiologically plausible bihemispheric cerebellar neuronal network model composed of spiking neurons as a platform to unveil roles and capacities of bihemispheric cerebellar architecture in the VOR motor learning.

Keiichiro Inagaki, Yutaka Hirata

Theory and Algorithms

Frontmatter

Modeling the Propensity Score with Statistical Learning

The progress of the ICT technology has produced data-sources that continuously generate datasets with different features and possibly with partial missing values. Such heterogeneity can be mended by integrating several processing blocks, but a unified method to extract conclusions from such heterogeneous datasets would bring consistent results with lower complexity. This paper proposes a flexible propensity score estimation method based on statistical learning for classification, and compared its performance against classical generalized linear methods.

Kenshi Uchihashi, Atsunori Kanemura

Analysis of the DNN-kWTA Network Model with Drifts in the Offset Voltages of Threshold Logic Units

The structure of the dual neural network-based (DNN) k-winner-take-all (kWTA) model is much simpler than that of other kWTA models. Its convergence time and capability under the perfect condition were reported. However, in the circuit implementation, the threshold levels of the threshold logic units (TLUs) in the DNN-kWTA model may have some drifts. This paper analyzes the DNN-kWTA model under the imperfect condition, where there are some drifts in the threshold level. We show that given that the inputs are uniformly distributed in the range of [0, 1], the probability that the DNN-kWTA model gives the correct output is greater than or equal to $$(1-2\varDelta )^n$$(1-2Δ)n, where $$\varDelta $$Δ is the maximum drift level. Besides, we derive the formulas for the average convergent time and the variance of the convergent time under the drift situation.

Ruibin Feng, Chi-Sing Leung, John Sum

Efficient Numerical Simulation of Neuron Models with Spatial Structure on Graphics Processing Units

Computer simulation of multi-compartment neuron models is difficult, because writing the computer program is tedious but complicated, and it requires sophisticated numerical methods to solve partial differential equations (PDEs) that describe the current flow in a neuron robustly. For this reason, dedicated simulation software such as NEURON and GENESIS have been used widely. However, these simulators do not support hardware acceleration using graphics processing units (GPUs). In this study, we implemented a conjugate gradient (CG) method to solve linear equations efficiently on a GPU in our own software. CG methods are known much faster and more efficient than the Gaussian elimination, when the matrix is huge and sparse. As a result, our software succeeded to carry out a simulation of Purkinje cells developed by De Schutter and Bower (1994) on a GPU. The GPU (Tesla K40c) version realized 3 times faster computation than that a single-threaded CPU version for 15 Purkinje cells.

Tsukasa Tsuyuki, Yuki Yamamoto, Tadashi Yamazaki

A Scalable Patch-Based Approach for RGB-D Face Recognition

This paper presents a novel approach for face recognition using low cost RGB-D cameras under challenging conditions. In particular, the proposed approach is based on salient points to extract local patches independently to the face pose. The classification is performed using a scalable sparse representation classification by an adaptive and dynamic dictionaries selection. The experimental results proved that the proposed algorithm achieves significant accuracy on three different RGB-D databases and competes with known approaches in the literature.

Nesrine Grati, Achraf Ben-Hamadou, Mohamed Hammami

Gaussian Processes Based Fusion of Multiple Data Sources for Automatic Identification of Geological Boundaries in Mining

Mining stratified ore deposits such as Banded Iron Formation (BIF) hosted iron ore deposits requires detailed knowledge of the location of orebody boundaries. In one Marra Mamba style deposit, the alluvial to bedded boundary only creates distinctive signatures when both the magnetic susceptibility logs and the downhole chemical assays are considered. Identifying where the ore to BIF boundary occurs with the NS3-NS4 stratigraphic boundary requires both natural gamma logs and chemical assays. These data sources have different downhole resolutions. This paper proposes a Gaussian Processes based method of probabilistically processing geophysical logs and chemical assays together. This method improves the classification of the alluvial to bedded boundary and allows the identification of concurring stratigraphic and mineralization boundaries. The results will help to automatically produce more accurate and objective geological models, significantly reducing the need for manual effort.

Katherine L. Silversides, Arman Melkumyan

Speaker Detection in Audio Stream via Probabilistic Prediction Using Generalized GEBI

This paper presents a method of speaker detection using probabilistic prediction for avoiding the tuning of thresholds to detect a speaker in an audio stream. We introduce g-GEBI (generalized GEBI) as a generalization of BI (Bayesian Inference) and GEBI (Gibbs-distribution-based Extended BI) to execute iterative detection of a speaker in audio stream uttered by more than one speaker. Then, we show a method of probabilistic prediction in multiclass classification to classify the results of speaker detection. By means of numerical experiments using recorded real speech data, we examine the properties and the effectiveness of the present method. Especially, we show that g-GEBI and g-BI (generalized BI) are more effective than the conventional BI and GEBI in incremental speaker detection task.

Koki Sakata, Shota Sakashita, Kazuya Matsuo, Shuichi Kurogi

Probabilistic Prediction for Text-Prompted Speaker Verification Capable of Accepting Spoken Words with the Same Meaning but Different Pronunciations

So far, we have presented a method of probabilistic prediction using GEBI (Gibbs-distribution based Bayesian inference) for flexible text-prompted speaker verification. For more flexible and practical verification, this paper presents a method of verification capable of accepting spoken words with the same meaning but different pronunciations. For example, Japanese language has different pronunciations for a digit, such as /yon/ and /shi/ for 4, /nana/ and /shichi/ for 7, which are usually uttered via unintentional selection, and then it is a practical problem in speech verification of words involving digits, such as ID numbers. With several assumptions, we present a modification of GEBI for dealing with such words. By means of numerical experiments using recorded real speech data, we examine the properties of the present method and show the validity and the effectiveness.

Shota Sakashita, Satoshi Takeguchi, Kazuya Matsuo, Shuichi Kurogi

Segment-Level Probabilistic Sequence Kernel Based Support Vector Machines for Classification of Varying Length Patterns of Speech

In this work we propose the segment-level probabilistic sequence kernel (SLPSK) as dynamic kernel to be used in support vector machine (SVM) for classification of varying length patterns of long duration speech represented as sets of feature vectors. SLPSK is built upon a set of Gaussian basis functions, where half of the basis functions contain class specific information while the other half implicates the common characteristics of all the speech utterances of all classes. The proposed kernel is computed between the pair of examples, by partitioning the speech signal into fixed number of segments and then matching the corresponding segments. We study the performance of the SVM-based classifiers using the proposed SLPSK using different pooling technique for speech emotion recognition and speaker identification and compare with that of the SVM-based classifiers using other kernels for varying length patterns.

Shikha Gupta, Veena Thenkanidiyoor, Dileep Aroor Dinesh

Attention Estimation for Input Switch in Scalable Multi-display Environments

Multi-Display Environments (MDEs) have become commonplace in office desks for editing and displaying different tasks, such as coding, searching, reading, and video-communicating. In this paper, we present a method of automatic switch for routing one input (including mouse/keyboard, touch pad, joystick, etc.) to different displays in scalable MDEs based on the user attention estimation. We set up an MDE in our office desk, in which each display is equipped with a webcam to capture the user’s face video for detecting if the user is looking at the display. We use Convolutional Neural Networks (CNNs) to learn the attention model from face videos with various poses, illuminations, and occlusions for achieving a high performance of attention estimation. Qualitative and quantitative experiments demonstrate the effectiveness and potential of the proposed approach. The results of the user study also shows that the participants deemed that the system is wonderful, useful, and friendly.

Xingyuan Bu, Mingtao Pei, Yunde Jia

Deep Dictionary Learning vs Deep Belief Network vs Stacked Autoencoder: An Empirical Analysis

A recent work introduced the concept of deep dictionary learning. The first level is a dictionary learning stage where the inputs are the training data and the outputs are the dictionary and learned coefficients. In subsequent levels of deep dictionary learning, the learned coefficients from the previous level acts as inputs. This is an unsupervised representation learning technique. In this work we empirically compare and contrast with similar deep representation learning techniques – deep belief network and stacked autoencoder. We delve into two aspects; the first one is the robustness of the learning tool in the presence of noise and the second one is the robustness with respect to variations in the number of training samples. The experiments have been carried out on several benchmark datasets. We find that the deep dictionary learning method is the most robust.

Vanika Singhal, Anupriya Gogna, Angshul Majumdar

Bi-directional LSTM Recurrent Neural Network for Chinese Word Segmentation

Recurrent neural network (RNN) has been broadly applied to natural language process (NLP) problems. This kind of neural network is designed for modeling sequential data and has been testified to be quite efficient in sequential tagging tasks. In this paper, we propose to use bi-directional RNN with long short-term memory (LSTM) units for Chinese word segmentation, which is a crucial task for modeling Chinese sentences and articles. Classical methods focus on designing and combining hand-craft features from context, whereas bi-directional LSTM network (BLSTM) does not need any prior knowledge or pre-designing, and is expert in creating hierarchical feature representation of contextual information from both directions. Experiment result shows that our approach gets state-of-the-art performance in word segmentation on both traditional Chinese datasets and simplified Chinese datasets.

Yushi Yao, Zheng Huang

Alternating Optimization Method Based on Nonnegative Matrix Factorizations for Deep Neural Networks

The backpropagation algorithm for calculating gradients has been widely used in computation of weights for deep neural networks (DNNs). This method requires derivatives of objective functions and has some difficulties finding appropriate parameters such as learning rate. In this paper, we propose a novel approach for computing weight matrices of fully-connected DNNs by using two types of semi-nonnegative matrix factorizations (semi-NMFs). In this method, optimization processes are performed by calculating weight matrices alternately, and backpropagation (BP) is not used. We also present a method to calculate stacked autoencoder using a NMF. The output results of the autoencoder are used as pre-training data for DNNs. The experimental results show that our method using three types of NMFs attains similar error rates to the conventional DNNs with BP.

Tetsuya Sakurai, Akira Imakura, Yuto Inoue, Yasunori Futamura

Fissionable Deep Neural Network

Model combination nearly always improves the performance of machine learning methods. Averaging the predictions of multi-model further decreases the error rate. In order to obtain multi high quality models more quickly, this article proposes a novel deep network architecture called “Fissionable Deep Neural Network”, abbreviated as FDNN. Instead of just adjusting the weights in a fixed topology network, FDNN contains multi branches with shared parameters and multi Softmax layers. During training, the model divides until to be multi models. FDNN not only can reduce computational cost, but also overcome the interference of convergence between branches and give an opportunity for the branches falling into a poor local optimal solution to re-learn. It improves the performance of neural network on supervised learning which is demonstrated on MNIST and CIFAR-10 datasets.

DongXu Tan, JunMin Wu, HuanXin Zheng, Yan Yin, YaXin Liu

A Structural Learning Method of Restricted Boltzmann Machine by Neuron Generation and Annihilation Algorithm

Restricted Boltzmann Machine (RBM) is a generative stochastic energy-based model of artificial neural network for unsupervised learning. The adaptive learning method that can discover the optimal number of hidden neurons according to the input space is important method in terms of the stability of energy as well as the computational cost although a traditional RBM model cannot change its network structure during learning phase. Moreover, we should consider the regularities in the sparse of network to extract explicit knowledge from the network because the trained network is often a black box. In this paper, we propose the combination method of adaptive and structural learning method of RBM with Forgetting that can discover the regularities in the trained network. We evaluated our proposed model on MNIST and CIFAR-10 datasets.

Shin Kamada, Takumi Ichimura

Semi-supervised Learning for Convolutional Neural Networks Using Mild Supervisory Signals

We propose a novel semi-supervised learning method for convolutional neural networks (CNNs). CNN is one of the most popular models for deep learning and its successes among various types of applications include image and speech recognition, image captioning, and the game of ‘go’. However, the requirement for a vast amount of labeled data for supervised learning in CNNs is a serious problem. Unsupervised learning, which uses the information of unlabeled data, might be key to addressing the problem, although it has not been investigated sufficiently in CNN regimes. The proposed method involves both supervised and unsupervised learning in identical feedforward networks, and enables seamless switching among them. We validated the method using an image recognition task. The results showed that learning using non-labeled data dramatically improves the efficiency of supervised learning.

Takashi Shinozaki

On the Singularity in Deep Neural Networks

In this paper, we analyze a deep neural network model from the viewpoint of singularities. First, we show that there exist a large number of critical points introduced by a hierarchical structure in the deep neural network as straight lines. Next, we derive sufficient conditions for the deep neural network having no critical points introduced by a hierarchical structure.

Tohru Nitta

A Deep Neural Network Architecture Using Dimensionality Reduction with Sparse Matrices

We present a new deep neural network architecture, motivated by sparse random matrix theory that uses a low-complexity embedding through a sparse matrix instead of a conventional stacked autoencoder. We regard autoencoders as an information-preserving dimensionality reduction method, similar to random projections in compressed sensing. Thus, exploiting recent theory on sparse matrices for dimensionality reduction, we demonstrate experimentally that classification performance does not deteriorate if the autoencoder is replaced with a computationally-efficient sparse dimensionality reduction matrix.

Wataru Matsumoto, Manabu Hagiwara, Petros T. Boufounos, Kunihiko Fukushima, Toshisada Mariyama, Zhao Xiongxin

Noisy Softplus: A Biology Inspired Activation Function

The Spiking Neural Network (SNN) has not achieved the recognition/classification performance of its non-spiking competitor, the Artificial Neural Network(ANN), particularly when used in deep neural networks. The mapping of a well-trained ANN to an SNN is a hot topic in this field, especially using spiking neurons with biological characteristics. This paper proposes a new biologically-inspired activation function, Noisy Softplus, which is well-matched to the response function of LIF (Leaky Integrate-and-Fire) neurons. A convolutional network (ConvNet) was trained on the MNIST database with Noisy Softplus units and converted to an SNN while maintaining a close classification accuracy. This result demonstrates the equivalent recognition capability of the more biologically-realistic SNNs and bring biological features to the activation units in ANNs.

Qian Liu, Steve Furber

Compressing Word Embeddings

Recent methods for learning vector space representations of words have succeeded in capturing fine-grained semantic and syntactic regularities using large-scale unlabelled text analysis. However, these representations typically consist of dense vectors that require a great deal of storage and cause the internal structure of the vector space to be opaque. A more ‘idealized’ representation of a vocabulary would be both compact and readily interpretable. With this goal, this paper first shows that Lloyd’s algorithm can compress the standard dense vector representation by a factor of 10 without much loss in performance. Then, using that compressed size as a ‘storage budget’, we describe a new GPU-friendly factorization procedure to obtain a representation which gains interpretability as a side-effect of being sparse and non-negative in each encoding dimension. Word similarity and word-analogy tests are used to demonstrate the effectiveness of the compressed representations obtained.

Martin Andrews

An Iterative Incremental Learning Algorithm for Complex-Valued Hopfield Associative Memory

This paper discusses a complex-valued Hopfield associative memory with an iterative incremental learning algorithm. The mathematical proofs derive that the weight matrix is approximated as a weight matrix by the complex-valued pseudo inverse algorithm. Furthermore, the minimum number of iterations for the learning sequence is defined with maintaining the network stability. From the result of simulation experiment in terms of memory capacity and noise tolerance, the proposed model has the superior ability than the model with a complex-valued pseudo inverse learning algorithm.

Naoki Masuyama, Chu Kiong Loo

LDA-Based Word Image Representation for Keyword Spotting on Historical Mongolian Documents

The original Bag-of-Visual-Words approach discards the spatial relations of the visual words. In this paper, a LDA-based topic model is adopted to obtain the semantic relations of visual words for each word image. Because the LDA-based topic model usually hurts retrieval performance when directly employs itself. Therefore, the LDA-based topic model is linearly combined with a visual language model for each word image in this study. After that, the basic query likelihood model is used for realizing the procedure of retrieval. The experimental results on our dataset show that the proposed LDA-based representation approach can efficiently and accurately attain to the aim of keyword spotting on a collection of historical Mongolian documents. Meanwhile, the proposed approach improves the performance significantly than the original BoVW approach.

Hongxi Wei, Guanglai Gao, Xiangdong Su

Solving the Vanishing Information Problem with Repeated Potential Mutual Information Maximization

The present paper shows how to solve the problem of vanishing information in potential mutual information maximization. We have previously developed a new information-theoretic method called “potential learning” which aims to extract the most important features through simplified information maximization. However, one of the major problems is that the potential effect diminishes considerably in the course of learning and it becomes impossible to take into account the potentiality in learning. To solve this problem, we here introduce repeated information maximization. To enhance the processes of information maximization, the method forces the potentiality to be assimilated in learning every time it becomes ineffective. The method was applied to the on-line article popularity data set to estimate the popularity of articles. To demonstrate the effectiveness of the method, the number of hidden neurons was made excessively large and set to 50. The results show that the potentiality information maximization could increase mutual information even with 50 hidden neurons, and lead to improved generalization performance. In addition, simplified representations could be obtained for better interpretation and generalization.

Ryotaro Kamimura

Self-organization on a Sphere with Application to Topological Ordering of Chinese Characters

We consider a case of self-organization in which a relatively small number N of data points is mapped on a larger number M of nodes. This is a reverse situation to a typical clustering problem when a node represents a center of the cluster of data points. In our case the objective is to have a Gaussian-like distribution of weights over nodes in the neighbourhood of the winner for a given stimulus. The fact that $$M\,>\,N$$M>N creates some problem with using learning schemes related to Gaussian Mixture Models. We also show how the objects, Chinese characters in our case, can be topologically ordered on a surface of a 3D sphere. A Chinese character is represented by an angular integral of the Radon Transform (aniRT) which is an RTS-invariant 1-D signature function of an image.

Andrew P. Papliński

A Spectrum Allocation Algorithm Based on Optimization and Protection in Cognitive Radio Networks

Cognitive radio network (CRN) is proposed to solve the problem of the scarce radio spectrum resources. In CRN, primary users (PUs) are allowed to lease out their unused spectrum sharing with cognitive users (CUs). In this paper, we propose a spectrum allocation algorithm based on CUs-demand and PUs-protection in CRN. Our objection is to make the allocated spectrum satisfy CUs demands as much as possible, avoiding the CUs interfering the PUs in the process of spectrum allocation. Simulation results indicate that this algorithm can improve the total spectrum reward, the satisfaction rate of CUs and the protection rate.

Jing Gao, Jianyu Lv, Xin Song

A Conjugate Gradient-Based Efficient Algorithm for Training Single-Hidden-Layer Neural Networks

A single hidden layer neural networks (SHLNNs) learning algorithm has been proposed which is called Extreme Learning Machine (ELM). It shows extremely faster than typical back propagation (BP) neural networks based on gradient descent method. However, it requires many more hidden neurons than BP neural networks to achieve assortive classification accuracy. This then leads more test time which plays an important role in practice. A novel learning algorithm (USA) for SHLNNs has been presented which updates the weights by using gradient method in the ELM framework. In this paper, we employ the conjugate gradient method to train the SHLNNs on the MNIST digit recognition problem. The simulated experiment demonstrates the better generalization and less required hidden neurons than the common ELM and USA.

Xiaoling Gong, Jian Wang, Yanjiang Wang, Jacek M. Zurada

The Ability of Learning Algorithms for Fuzzy Inference Systems Using Vector Quantization

Many studies on learning of fuzzy inference systems have been made. Specifically, it is known that learning methods using VQ (Vector Quantization) and SDM (Steepest Descend Method) are superior to other methods. We already proposed new learning methods iterating VQ and SDM. In their learning methods, VQ is used only in determination of parameters for the antecedent part of fuzzy rules. In order to improve them, we added the method determining of parameters for the consequent part of fuzzy rules to processing of VQ and SDM. That is, we proposed a learning method composed of three stages as VQ, GIM(Generalized Inverse Matrix) and SDM in the previous paper. In this paper, the ability of the proposed method is compared with other ones using VQ. As a result, it is shown that the proposed method outperforms conventional ones using VQ in terms of accuracy and the number of rules.

Hirofumi Miyajima, Noritaka Shigei, Hiromi Miyajima

An Improved Multi-strategy Ensemble Artificial Bee Colony Algorithm with Neighborhood Search

Artificial bee colony (ABC) algorithm has been shown its good performance over many optimization problems. Recently, a multi-strategy ensemble ABC (MEABC) algorithm was proposed which employed three distinct solution search strategies. Although its such mechanism works well, it may run the risk of causing the problem of premature convergence when solving complex optimization problems. Hence, we present an improved version by integrating the neighborhood search operator of which object is to perturb the global best food source for better balancing the exploration and exploitation. Experiments are conducted on a set of 22 well-known benchmark functions, and the results show that both of the quality of final results and convergence speed can be improved.

Xinyu Zhou, Mingwen Wang, Jianyi Wan, Jiali Zuo

Gender-Specific Classifiers in Phoneme Recognition and Academic Emotion Detection

Gender-specific classifiers are shown to outperform general classifiers. In calibrated experiments designed to demonstrate this, two sets of data were used to build male-specific and female-specific classifiers. The first dataset is used to predict vowel phonemes based on speech signals, and the second dataset is used to predict negative emotions based on brainwave (EEG) signals. A Multi-Layered-Perceptron (MLP) is first trained as a general classifier, where all data from both male and female users are combined. This general classifier recognizes vowel phonemes with a baseline accuracy of 91.09 %, while that for EEG signals has an average baseline accuracy of 58.70 %. The experiments show that the performance significantly improves when the classifiers are trained to be gender-specific – that is, there is a separate classifier for male users, and a separate classifier for female users. For the vowel phoneme recognition dataset, the average accuracy increases to 94.20 % and 95.60 %, for male only users and female-only users, respectively. As for the EEG dataset, the accuracy increases to 65.33 % for male-only users and to 70.50 % for female-only users. Performance rates using recall and precision show the same trend. A further probe is done using SOM to visualize the distribution of the sub-clusters among male and female users.

Arnulfo Azcarraga, Arces Talavera, Judith Azcarraga

Local Invariance Representation Learning Algorithm with Multi-layer Extreme Learning Machine

Multi-layer extreme learning machine (ML-ELM) is a stacked extreme learning machine based auto-encoding (ELM-AE). It provides an effective solution for deep feature extraction with higher training efficiency. To enhance the local-input invariance of feature extraction, we propose a contractive multi-layer extreme learning machine (C-ML-ELM) by adding a penalty term in the optimization function to minimize derivative of output to input at each hidden layer. In this way, the extracted feature is supposed to keep consecutiveness attribution of an image. The experiments have been done on MNIST handwriting dataset and face expression dataset CAFÉ. The results show that it outperforms several state-of-art classification algorithms with less error and higher training efficiency.

Xibin Jia, Xiaobo Li, Hua Du, Bir Bhanu

Two-Dimensional Soft Linear Discriminant Projection for Robust Image Feature Extraction and Recognition

In this study, we propose a Robust Soft Linear Discriminant Projection (RS-LDP) algorithm for extracting two-dimensional (2D) image features for recognition. RS-LDP is based on the soft label linear discriminant analysis (SL-LDA) that is shown to be effective for semi-supervised feature learning, but SLDA works in the vector space and thus extract one-dimensional (1D) features directly, so it has to convert the two-dimensional (2D) image matrices into the 1D vectorized representations in a high-dimensional space when dealing with images. But such transformation usually destroys the intrinsic topology structures of the images pixels and thus loses certain important information, which may result in degraded performance. Compared with SL-LDA for representation, our RS-LDP can effectively preserve the topology structures among image pixels, and more importantly it would be more efficient due to the matrix representations. Extensive simulations on real-world image datasets show that our proposed RS-LDP can deliver enhanced performance over other state-of-the-arts for recognition.

Yu Tang, Zhao Zhang, Weiming Jiang

Asymmetric Synaptic Connections in Z(2) Gauge Neural Network

We consider Z(2) gauge neural network which involves neuron variables $$S_i (=\pm 1)$$Si(=±1) and synaptic connection (gauge) variables $$J_{ij} (=\pm 1)$$Jij(=±1). Its energy consists of the Hopfield term $$-c_1S_iJ_{ij}S_j$$-c1SiJijSj and the reverberation term $$-c_2J_{ij}J_{jk}J_{ki}$$-c2JijJjkJki for signal propagation on a closed loop. The model of symmetric couplings $$J_{ij}=J_{ji}$$Jij=Jji has been studied; its phase diagram in the $$c_2-c_1$$c2-c1 plane exhibits Higgs, Coulomb and confinement phases, each of which is characterized by the ability of learning and/or recalling patterns. In this paper, we consider the model of asymmetric coupling ($$J_{ij}$$Jij and $$J_{ji}$$Jji are independent), and examine the effect of asymmetry on a partially connected random network.

Atsutomo Murai, Tetsuo Matsui

SOMphony: Visualizing Symphonies Using Self Organizing Maps

Symphonies are musical compositions played by a full orchestra which have evolved in style since the 16th Century. Self-Organizing Maps (SOM) are shown to be useful in visualizing symphonies as a musical trajectory across the nodes in a trained map. This allows for some insights about the relationships and influences between and among composers in terms of their composition styles, and how the symphonic compositions have evolved over the years from one major music period to the next. The research focuses on Self Organizing Maps that are trained using 1-second music segments extracted from 45 different symphonies, from 15 different composers, with 3 composers from each of the 5 major musical periods. The trained SOM is further processed by doing a k-means clustering of the node vectors, which then allows for the quantitative comparison of music trajectories between symphonies of the same composer, between symphonies of different composers of the same music period, and between composers from different music periods.

Arnulfo Azcarraga, Fritz Kevin Flores

Online EM for the Normalized Gaussian Network with Weight-Time-Dependent Updates

In this paper, we propose a weight-time-dependent (WTD) update approach for an online EM algorithm applied to the Normalized Gaussian network (NGnet). WTD aims to improve a recently proposed weight-dependent (WD) update approach by Celaya and Agostini. First, we discuss the derivation of WD from an older time-dependent (TD) update approach. Then, we consider additional aspects to improve WD, and by including them we derive the new WTD approach from TD. The difference between WD and WTD is discussed, and some experiments are conducted to demonstrate the effectiveness of the proposed approach. WTD succeeds in improving the learning performance for a function approximation task with balanced and dynamic data distributions.

Jana Backhus, Ichigaku Takigawa, Hideyuki Imai, Mineichi Kudo, Masanori Sugimoto

Learning Phrase Representations Based on Word and Character Embeddings

Most phrase embedding methods consider a phrase as a basic term and learn embeddings according to phrases’ external contexts, ignoring the internal structures of words and characters. There are some languages such as Chinese, a phrase is usually composed of several words or characters and contains rich internal information. The semantic meaning of a phrase is also related to the meanings of its composing words or characters. Therefore, we take Chinese for example, and propose a joint words and characters embedding model for learning phrase representation. In order to disambiguate the word and character and address the issue of non-compositional phrases, we present multiple-prototype word and character embeddings and an effective phrase selection method. We evaluate the effectiveness of the proposed model on phrase similarities computation and analogical reasoning. The empirical result shows that our model outperforms other baseline methods which ignore internal word and character information.

Jiangping Huang, Donghong Ji, Shuxin Yao, Wenzhi Huang, Bo Chen

A Mobile-Based Obstacle Detection Method: Application to the Assistance of Visually Impaired People

Visual impairments suffer many difficulties when they navigate from one place to another in their daily life. The biggest problem is obstacle detection. In this work, we propose a new smartphone-based method for obstacle detection. We aim to detect static and dynamic obstacles in unknown environments while offering maximum flexibility to the user and using the least expensive equipment possible. Detecting obstacles is based on the analysis of different regions of video frames and using a new decision algorithm. The analysis uses prediction model for each region that generated by a supervised learning process. The user is notified about the existing of an obstacle by alert message. The efficiency of the work is measured by many experiments studies on different complex scenes. It records low false alarm rate in the range of [0.2 % to 11 %], and high accuracy in the range of [86 % to 94 %].

Manal Abdulaziz Alshehri, Salma Kammoun Jarraya, Hanene Ben-Abdallah

t-SNE Based Visualisation and Clustering of Geological Domain

Identification of geological domains and their boundaries plays a vital role in the estimation of mineral resources. Geologists are often interested in exploratory data analysis and visualization of geological data in two or three dimensions in order to detect quality issues or to generate new hypotheses. We compare PCA and some other linear and non-linear methods with a newer method, t-Distributed Stochastic Neighbor Embedding (t-SNE) for the visualization of large geochemical assay datasets. The t-SNE based reduced dimensions can then be used with clustering algorithm to extract well clustered geological regions using exploration and production datasets. Significant differences between the nonlinear method t-SNE and the state of the art methods were observed in two dimensional target spaces.

Mehala Balamurali, Arman Melkumyan

Data-Based Optimal Tracking Control of Nonaffine Nonlinear Discrete-Time Systems

The optimal tracking control problem of nonaffine nonlinear discrete-time systems is considered in this paper. The problem relies on the solution of the so-called tracking Hamilton-Jacobi-Bellman equation, which is extremely difficult to be solved even for simple systems. To overcome this difficulty, the data-based Q-learning algorithm is proposed by learning the optimal tracking control policy from data of the practical system. For its implementation purpose, the critic-only neural network structure is developed, where only critic neural network is required to estimate the Q-function and the least-square scheme is employed to update the weight of neural network.

Biao Luo, Derong Liu, Tingwen Huang, Chao Li

Time Series Classification Based on Multi-codebook Important Time Subsequence Approximation Algorithm

This paper proposes a multi-codebook important time subsequence approximation (MCITSA) algorithm for time series classification. MCITSA generates a codebook using important time subsequences for each class based on the difference of categories. In this way, each codebook contains the class information itself. To predict the class label of an unseen time series, MCITSA needs to compare the similarities between important time subsequences extracted from the unseen time series and codewords of each class. Experimental results on time series datasets demonstrate that MCITSA is more powerful than PVQA in classifying time series.

Zhiwei Tao, Li Zhang, Bangjun Wang, Fanzhang Li

Performance Improvement via Bagging in Ensemble Prediction of Chaotic Time Series Using Similarity of Attractors and LOOCV Predictable Horizon

Recently, we have presented a method of ensemble prediction of chaotic time series. The method employs strong learners capable of making predictions with small error and usual ensemble mean does not work for long term prediction owing to the long term unpredictability of chaotic time series. Thus, the method uses similarity of attractors to select plausible predictions from original predictions generated by strong leaners, and then calculates LOOCV (leave-one-out cross-validation) measure to estimate predictable horizons. Finally, it provides representative prediction and an estimation of the predictable horizon. We have used CAN2s (competitive associative nets) for learning piecewise linear approximation of nonlinear function as strong learners in the previous study, and this paper employs bagging of them to improve the performance, and shows the validity and the effectiveness of the method.

Mitsuki Toidani, Kazuya Matsuo, Shuichi Kurogi

A Review of EEG Signal Simulation Methods

This paper describes EEG signal simulation methods. Three main methods have been included in this study: Markov Process Amplitude (MPA), Artificial Neural Network (ANN), and Autoregressive (AR) models. Each method is described procedurally, along with mathematical expressions. By the end of the description of each method, the limitations and benefits are described in comparison with other methods. MPA comprises of three variations; first-order MPA, nonlinear MPA, and adaptive MPA. ANN consists of two variations; feed forward back-propagation NN and multilayer feed forward with error back-propagation NN with embedded driving signal. AR model based filtering has been considered with its variation, genetic algorithm based on autoregressive moving average (ARMA) filtering.

Muhammad Izhan Noorzi, Ibrahima Faye

A New Blind Image Quality Assessment Based on Pairwise

Recently, the algorithms of general purpose blind image quality assessment (BIQA) have been an important research area in the field of image processing, but the previous approaches usually depend on human scores image for training and using the regression methods to predict the image quality. In this paper, we first apply the full-reference image quality measure to obtain the image quality scores for training to let our algorithm independent of the judgment of human. Then, we abstract features using an NSS model of the image DCT coefficient which is indicative of perceptual quality, and subsequently, we import Pairwise approach of Learning to rank (machine-learned ranking) to predict the perceptual scores of image quality. Our algorithm is tested on LIVE II and CSIQ database and it is proved to perform highly correlate with human judgment of image quality, and better than the popular SSIM index and competitive with the state-of-the-art BIQA algorithms.

Jianbin Jiang, Yue Zhou, Liming He

Self-organizing Maps as Feature Detectors for Supervised Neural Network Pattern Recognition

Convolutional neural network (CNN)-based works show that learned features, rather than handpicked features, produce more desirable performance in pattern recognition. This learning approach is based on higher organisms visual system which are developed based on the input environment. However, the feature detectors of CNN are trained using an error-correcting teacher as opposed to the natural competition to build node connections. As such, a neural network model using self-organizing map (SOM) as feature detector is proposed in this work. As proof of concept, the handwritten digits dataset is used to test the performance of the proposed architecture. The size of the feature detector as well as the different arrangement of receptive fields are considered to benchmark the performance of the proposed network. The performance for the proposed architecture achieved comparable performance to vanilla MLP, being 96.93 % using 4$$\times $$×4 SOM and six receptive field regions.

Macario O. Cordel, Arren Matthew C. Antioquia, Arnulfo P. Azcarraga

A Review of Electroencephalogram-Based Analysis and Classification Frameworks for Dyslexia

Dyslexia is a hidden learning disability that causes difficulties in reading and writing despite average intelligence. Electroencephalogram (EEG) is one of the upcoming methods being researched for identifying unique brain activation patterns in dyslexics. This paper examines pros and cons of existing EEG-based analysis and classification frameworks for dyslexia and recommends optimizations through the findings to assist future research.

Harshani Perera, Mohd Fairuz Shiratuddin, Kok Wai Wong

Rule-Based Grass Biomass Classification for Roadside Fire Risk Assessment

Roadside grass fire is a major hazard to the security of drivers and vehicles. However, automatic assessment of roadside grass fire risk has not been fully investigated. This paper presents an approach, for the first time to our best knowledge, that automatically estimates and classifies grass biomass for determining the fire risk level of roadside grasses from video frames. A major novelty is automatic measurement of grass coverage and height for predicting the biomass. For a sampling grass region, the approach performs two-level grass segmentation using class-specific neural networks. The brown grass coverage is then calculated and an algorithm is proposed that uses continuously connected vertical grass pixels to estimate the grass height. Based on brown grass coverage and grass height, a set of threshold based rules are designed to classify grasses into low, medium or high risk. Experiments on a challenging real-world dataset demonstrate promising results of our approach.

Ligang Zhang, Brijesh Verma

Efficient Recognition of Attentional Bias Using EEG Data and the NeuCube Evolving Spatio-Temporal Data Machine

Modelling of dynamic brain activity for better understanding of human decision making processes becomes an important task in many areas of study. Inspired by importance of the attentional bias principle in human choice behaviour, we proposed a Spiking Neural Network (SNN) model for efficient recognition of attentional bias. The model is based on the evolving spatio-temporal data machine NeuCube. The proposed model is tested on a case study experimental EEG data collected from a group of subjects exemplified here on a group of moderate drinkers when they were presented by different product features (in this case different features of drinks). The results showed a very high accuracy of discriminating attentional bias to non-target objects and their features when compared with a poor performance of traditional machine learning methods. Potential applications in neuromarketing and cognitive studies are also discussed.

Zohreh Gholami Doborjeh, Maryam Gholami Doborjeh, Nikola Kasabov

Backmatter

Weitere Informationen

Premium Partner

Neuer Inhalt

BranchenIndex Online

Die B2B-Firmensuche für Industrie und Wirtschaft: Kostenfrei in Firmenprofilen nach Lieferanten, Herstellern, Dienstleistern und Händlern recherchieren.

Whitepaper

- ANZEIGE -

Best Practices für die Mitarbeiter-Partizipation in der Produktentwicklung

Unternehmen haben das Innovationspotenzial der eigenen Mitarbeiter auch außerhalb der F&E-Abteilung erkannt. Viele Initiativen zur Partizipation scheitern in der Praxis jedoch häufig. Lesen Sie hier  - basierend auf einer qualitativ-explorativen Expertenstudie - mehr über die wesentlichen Problemfelder der mitarbeiterzentrierten Produktentwicklung und profitieren Sie von konkreten Handlungsempfehlungen aus der Praxis.
Jetzt gratis downloaden!

Bildnachweise