A method for encoding detection and bit rate classification of AMR-coded speech is proposed. For each texture frame, 184 features consisting of the short-term and long-term temporal statistics of speech parameters are extracted, which can effectively measure the amount of distortion due to AMR. The deep neural network then classifies the bit rate of speech after analyzing the extracted features. It is confirmed that the proposed features provide better performance than the conventional spectral features designed for bit rate classification of coded audio.

In this study, toward the development of smartphone-based monitoring system for life logging, we collect over 1,400 hours of data by recording including both the outdoor and indoor daily activities of 19 subjects, under practical conditions with a smartphone and a small camera. We then construct a huge human activity database which consists of an environmental sound signal, triaxial acceleration signals and manually annotated activity tags. Using our constructed database, we evaluate the activity recognition performance of deep neural networks (DNNs), which have achieved great performance in various fields, and apply DNN-based adaptation techniques to improve the performance with only a small amount of subject-specific training data. We experimentally demonstrate that; 1) the use of multi-modal signal, including environmental sound and triaxial acceleration signals with a DNN is effective for the improvement of activity recognition performance, 2) the DNN can discriminate specified activities from a mixture of ambiguous activities, and 3) DNN-based adaptation methods are effective even if only a small amount of subject-specific training data is available.

In this paper, we propose a deep learning based model for classifying speech-acts using a convolutional neural network (CNN). The model uses some bigram features including parts-of-speech (POS) tags and dependency-relation bigrams, which represent syntactic structural information in utterances. Previous classification approaches using CNN have commonly exploited word embeddings using morpheme unigrams. However, the proposed model first extracts two different bigram features that well reflect the syntactic structure of utterances and then represents them as a vector representation using a word embedding technique. As a result, the proposed model using bigram embeddings achieves an accuracy of 89.05%. Furthermore, the accuracy of this model is relatively 2.8% higher than that of competitive models in previous studies.

This letter describes the development and implementation of the lane detection system accelerated by the neuromorphic hardware. Because the neuromorphic hardware has inherently parallel nature and has constant output latency regardless the size of the knowledge, the proposed lane detection system can recognize various types of lanes quickly and efficiently. Experimental results using the road images obtained in the actual driving environments showed that white and yellow lanes could be detected with an accuracy of more than 94 percent.

In industrial control processes, control performance influences the quality of products and utilization efficiency of energy; hence, the controller is necessarily designed according to user-desired control performance. Ideal control performance requires fast response for transient state and maintaining user-specified control performance for steady state. Hence, an algorithm to tune controller parameters to match the requirements for transient state and steady state is proposed. Considering the partial learning ability of the cerebellar model articulation controller (CMAC) neural network, it is utilized as a “tuner” of controller parameters in this study, since then the controller parameters can be tuned in both transient and steady states. Moreover, the fictitious reference iterative tuning (FRIT) algorithm is combined with CMAC in order to avoid problems, which may be caused by system modeling error and by using only a set of closed-loop data, the desired controller can be calculated in an off-line manner. In addition, the controller selected is a proportional-integral-derivative (PID) controller. Finally, the effectiveness of the proposed method is numerically verified by using some simulation and experimental examples.

In an IoT system, neural networks have the potential to perform advanced information processing in various environments. To clarify this, the robustness of a restricted Boltzmann machine (RBM) used for deep neural networks, such as a deep belief network (DBN), was studied in this paper. Even if memory or logic errors occurred in the circuit operating in the RBM while pre-training the DBN, they did not affect the identification rate of the DBN, showing the robustness of the RBM. In addition, robustness against soft errors was evaluated. The soft errors had almost no influence on the RBM unless they were as large as 1012 times or more in the 50-nm CMOS process.

Accurate stereo matching remains a challenging problem in case of weakly-textured areas, discontinuities and occlusions. In this letter, a novel stereo matching method, consisting of leveraging feature ensemble network to compute matching cost, error detection network to predict outliers and priority-based occlusion disambiguation for refinement, is presented. Experiments on the Middlebury benchmark demonstrate that the proposed method yields competitive results against the state-of-the-art algorithms.

Various types of classifiers and feature extraction methods for acoustic scene classification have been recently proposed in the IEEE Detection and Classification of Acoustic Scenes and Events (DCASE) 2016 Challenge Task 1. The results of the final evaluation, however, have shown that even top 10 ranked teams, showed extremely low accuracy performance in particular class pairs with similar sounds. Due to such sound classes being difficult to distinguish even by human ears, the conventional deep learning based feature extraction methods, as used by most DCASE participating teams, are considered facing performance limitations. To address the low performance problem in similar class pair cases, this letter proposes to employ a recurrent neural network (RNN) based source separation for each class prior to the classification step. Based on the fact that the system can effectively extract trained sound components using the RNN structure, the mid-layer of the RNN can be considered to capture discriminative information of the trained class. Therefore, this letter proposes to use this mid-layer information as novel discriminative features. The proposed feature shows an average classification rate improvement of 2.3% compared to the conventional method, which uses additional classifiers for the similar class pair issue.

The latest deep hashing methods perform hash codes learning and image feature learning simultaneously by using pairwise or triplet labels. However, generating all possible pairwise or triplet labels from the training dataset can quickly become intractable, where the majority of those samples may produce small costs, resulting in slow convergence. In this letter, we propose a novel deep discriminative supervised hashing method, called DDSH, which directly learns hash codes based on a new combined loss function. Compared to previous methods, our method can take full advantages of the annotated data in terms of pairwise similarity and image identities. Extensive experiments on standard benchmarks demonstrate that our method preserves the instance-level similarity and outperforms state-of-the-art deep hashing methods in the image retrieval application. Remarkably, our 16-bits binary representation can surpass the performance of existing 48-bits binary representation, which demonstrates that our method can effectively improve the speed and precision of large scale image retrieval systems.

This paper presents convex filter networks that are obtained from extensions of morphological filters. The proposed filter network consists of a convex and concave filter that are extensions of the dilation and erosion of mathematical morphology with the maxout activation function. Maxout can approximate arbitrary convex functions as piecewise linear functions, including the max function. The class of the convex function hence includes the morphological dilation and can be trained for specific image processing tasks. In this paper, the closing filter is extended to a convex-concave filter network with maxout. The convex-concave filter is trained by the stochastic gradient method for noise and mask removal. The examples of noise and mask removal show that the convex-concave filter can obtain a recovered image, whose quality is comparable to inpainting by using the total variation minimization with reduced computational cost without mask information of the corrupted pixels.

There is increasing interest in quantum computing, because of its enormous computing potential. A small number of powerful quantum algorithms have been proposed to date; however, the development of new quantum algorithms for practical use remains essential. Parallel computing with a neural network has successfully realized certain unique functions such as learning and recognition; therefore, the introduction of certain neural computing techniques into quantum computing to enlarge the quantum computing application field is worthwhile. In this paper, a novel quantum associative memory (QuAM) is proposed, which is achieved with a quantum neural network by employing adiabatic Hamiltonian evolution. The memorization and retrieval procedures are inspired by the concept of associative memory realized with an artificial neural network. To study the detailed dynamics of our QuAM, we examine two types of Hamiltonians for pattern memorization. The first is a Hamiltonian having diagonal elements, which is known as an Ising Hamiltonian and which is similar to the cost function of a Hopfield network. The second is a Hamiltonian having non-diagonal elements, which is known as a neuro-inspired Hamiltonian and which is based on interactions between qubits. Numerical simulations indicate that the proposed methods for pattern memorization and retrieval work well with both types of Hamiltonians. Further, both Hamiltonians yield almost identical performance, although their retrieval properties differ. The QuAM exhibits new and unique features, such as a large memory capacity, which differs from a conventional neural associative memory.

Coaxial connectors are extensively used in electrical systems and the degradation of the connector can alter the signal that is being transmitted and leads to faults, which is one of the major causes of low communication quality. In this work, the failure features caused by the degraded connector contact surface were studied. The relationship between the DC resistance and decreased real contact areas was given. Considering the inductance properties and capacitive coupling at high frequencies, the impedance characteristics of the degraded connector were discussed. Based on the transmission line theory and experimental measurement, an equivalent lump circuit of the coaxial connector was developed. For the degraded contact surface, the capacitance was analyzed, and the frequency effect was investigated. According to the high frequency characteristics of the degraded connector, a fault detection and location method for coaxial connectors in RF system was developed using a neural network method. For connectors suffering from different levels of pollution, their impedance modulus varies continuously. Considering the range of the connector's impedance parameters, the fault modes were determined. Based on the scattering parameter simulation of a RF receiver front-end circuit, the S11 and S21 parameters were obtained as feature parameters and Monte Carlo simulations were conducted to generate training and testing samples. Based on the BP neural network algorithm, the fault modes were classified and the results show the diagnosis accuracy was 97.33%.

Detection of cavities in X-ray astronomical images has become a field of interest, since the flourishing studies on black holes and the Active Galactic Nuclei (AGN). In this paper, an approach is proposed to detect cavities in X-ray astronomical images using our newly designed Granular Convolutional Neural Network (GCNN) based classifiers. The raw data are firstly preprocessed to obtain images of the observed objects, i.e., galaxies or galaxy clusters. In each image, pixels are classified into three categories, (1) the faint backgrounds (BKG), (2) the cavity regions (CAV), and (3) the bright central gas regions (CNT). And the sample sets are then generated by dividing large images into subimages with a window size according to the cavities' scale. Since the number of BKG samples are far more than the other types, to achieve balanced training sets, samples from the major class are split into subsets, i.e., granule. Then a group of three-convolutional-layer granular CNN networks without subsampling layers are designed as the classifiers, and trained with the labeled granular sample sets. Finally, the trained GCNN classifiers are applied to new observations, so as to estimate the cavity regions with a voting strategy and locate them with elliptical profiles on the raw observation images. Experiments and applications of our approach are demonstrated on 40 X-ray astronomical observations retrieved from chandra Data Archive (CDA). Comparisons among our approach, the β-model fitting and the Unsharp Masking (UM) methods were also performed, which prove our approach was more accurate and robust.

In this paper, a novel approach is proposed for stereo vision-based ground plane detection at superpixel-level, which is implemented by employing a Disparity Texture Map in a convolution neural network architecture. In particular, the Disparity Texture Map is calculated with a new Local Disparity Texture Descriptor (LDTD). The experimental results demonstrate our superior performance in KITTI dataset.

Conventional feature-rich parsers based on manually tuned features have achieved state-of-the-art performance. However, these parsers are not good at handling long-term dependencies using only the clues captured by a prepared feature template. On the other hand, recurrent neural network (RNN)-based parsers can encode unbounded history information effectively, but they perform not well for small tree structures, especially when low-frequency words are involved, and they cannot use prior linguistic knowledge. In this paper, we propose a simple but effective framework to combine the merits of feature-rich transition-based parsers and RNNs. Specifically, the proposed framework incorporates RNN-based scores into the feature template used by a feature-rich parser. On English WSJ treebank and SPMRL 2014 German treebank, our framework achieves state-of-the-art performance (91.56 F-score for English and 83.06 F-score for German), without requiring any additional unlabeled data.

Recent acoustic event classification research has focused on training suitable filters to represent acoustic events. However, due to limited availability of target event databases and linearity of conventional filters, there is still room for improving performance. By exploiting the non-linear modeling of deep neural networks (DNNs) and their ability to learn beyond pre-trained environments, this letter proposes a DNN-based feature extraction scheme for the classification of acoustic events. The effectiveness and robustness to noise of the proposed method are demonstrated using a database of indoor surveillance environments.

Recently, numerous methods have been proposed to tackle the problem of fine-grained image classification. However, rare of them focus on the pre-processing step of image alignment. In this paper, we propose a new pre-processing method with the aim of reducing the variance of objects among the same class. As a result, the variance of objects between different classes will be more significant. The proposed approach consists of four procedures. The “parts” of the objects are firstly located. After that, the rotation angle and the bounding box could be obtained based on the spatial relationship of the “parts”. Finally, all the images are resized to similar sizes. The objects in the images possess the properties of translation, scale and rotation invariance after processed by the proposed method. Experiments on the CUB-200-2011 and CUB-200-2010 datasets have demonstrated that the proposed method could boost the recognition performance by serving as a pre-processing step of several popular classification algorithms.

This paper investigates the effect of noises added to hidden units of AutoEncoders linked to multilayer perceptrons. It is shown that internal representation of learned features emerges and sparsity of hidden units increases when independent Gaussian noises are added to inputs of hidden units during the deep network training. It is also shown that the weights that connect the contaminated hidden units with the next layer have smaller values and outputs of hidden units tend to be more definite (0 or 1). This is expected to improve the generalization ability of the network through this automatic structuration by adding the noises. This network structuration was confirmed by experiments for MNIST digits classification via a deep neural network model.

Quaternionic neural networks are extensions of neural networks using quaternion algebra. 3-D and 4-D quaternionic MLPs have been studied. 3-D quaternionic neural networks are useful for handling 3-D objects, such as Euclidean transformation. As for Hopfield neural networks, only 4-D quaternionic Hopfield neural networks (QHNNs) have been studied. In this work, we propose the 3-D QHNNs. Moreover, we define the energy, and prove that it converges.

Due to the increase of outsourcing by IC vendors, we face a serious risk that malicious third-party vendors insert hardware Trojans very easily into their IC products. However, detecting hardware Trojans is very difficult because today's ICs are huge and complex. In this paper, we propose a hardware-Trojan classification method for gate-level netlists to identify hardware-Trojan infected nets (or Trojan nets) using a support vector machine (SVM) or a neural network (NN). At first, we extract the five hardware-Trojan features from each net in a netlist. These feature values are complicated so that we cannot give the simple and fixed threshold values to them. Hence we secondly represent them to be a five-dimensional vector and learn them by using SVM or NN. Finally, we can successfully classify all the nets in an unknown netlist into Trojan ones and normal ones based on the learned classifiers. We have applied our machine-learning-based hardware-Trojan classification method to Trust-HUB benchmarks. The results demonstrate that our method increases the true positive rate compared to the existing state-of-the-art results in most of the cases. In some cases, our method can achieve the true positive rate of 100%, which shows that all the Trojan nets in an unknown netlist are completely detected by our method.