The recent development of deep learning-based generative models has sharply intensified the interest in data synthesis and its applications. Data synthesis takes on an added importance especially for some pattern recognition tasks in which some classes of data are rare and difficult to collect. In an iris dataset, for instance, the minority class samples include images of eyes with glasses, oversized or undersized pupils, misaligned iris locations, and iris occluded or contaminated by eyelids, eyelashes, or lighting reflections. Such class-imbalanced datasets often result in biased classification performance. Generative adversarial networks (GANs) are one of the most promising frameworks that learn to generate synthetic data through a two-player minimax game between a generator and a discriminator. In this paper, we utilized the state-of-the-art conditional Wasserstein generative adversarial network with gradient penalty (CWGAN-GP) for generating the minority class of iris images which saves huge amount of cost of human labors for rare data collection. With our model, the researcher can generate as many iris images of rare cases as they want and it helps to develop any deep learning algorithm whenever large size of dataset is needed.

Relation extraction is one of the key basic tasks in natural language processing in which distant supervision is widely used for obtaining large-scale labeled data without expensive labor cost. However, the automatically generated data contains massive noise because of the wrong labeling problem in distant supervision. To address this problem, the existing research work mainly focuses on removing sentence-level noise with various sentence selection strategies, which however could be incompetent for disposing word-level noise. In this paper, we propose a novel neural framework considering both intra-sentence and inter-sentence relevance to deal with word-level and sentence-level noise from distant supervision, which is denoted as Sentence-Related Gated Piecewise Convolutional Neural Networks (SR-GPCNN). Specifically, 1) a gate mechanism with multi-head self-attention is adopted to reduce word-level noise inside sentences; 2) a soft-label strategy is utilized to alleviate wrong-labeling propagation problem; and 3) a sentence-related selection model is designed to filter sentence-level noise further. The extensive experimental results on NYT dataset demonstrate that our approach filters word-level and sentence-level noise effectively, thus significantly outperforms all the baseline models in terms of both AUC and top-n precision metrics.

Gait recognition distinguishes one individual from others according to the natural patterns of human gaits. Gait recognition is a challenging signal processing technology for biometric identification due to the ambiguity of contours and the complex feature extraction procedure. In this work, we proposed a new model - the convolutional neural network (CNN) joint attention mechanism (CJAM) - to classify the gait sequences and conduct person identification using the CASIA-A and CASIA-B gait datasets. The CNN model has the ability to extract gait features, and the attention mechanism continuously focuses on the most discriminative area to achieve person identification. We present a comprehensive transformation from gait image preprocessing to final identification. The results from 12 experiments show that the new attention model leads to a lower error rate than others. The CJAM model improved the 3D-CNN, CNN-LSTM (long short-term memory), and the simple CNN by 8.44%, 2.94% and 1.45%, respectively.

This paper proposes two high-fidelity and computationally efficient neural voice conversion (VC) methods based on a direct waveform modification using spectral differentials. The conventional spectral-differential VC method with a minimum-phase filter achieves high-quality conversion for narrow-band (16 kHz-sampled) VC but requires heavy computational cost in filtering. This is because the minimum phase obtained using a fixed lifter of the Hilbert transform often results in a long-tap filter. Furthermore, when we extend the method to full-band (48 kHz-sampled) VC, the computational cost is heavy due to increased sampling points, and the converted-speech quality degrades due to large fluctuations in the high-frequency band. To construct a short-tap filter, we propose a lifter-training method for data-driven phase reconstruction that trains a lifter of the Hilbert transform by taking into account filter truncation. We also propose a frequency-band-wise modeling method based on sub-band multi-rate signal processing (sub-band modeling method) for full-band VC. It enhances the computational efficiency by reducing sampling points of signals converted with filtering and improves converted-speech quality by modeling only the low-frequency band. We conducted several objective and subjective evaluations to investigate the effectiveness of the proposed methods through implementation of the real-time, online, full-band VC system we developed, which is based on the proposed methods. The results indicate that 1) the proposed lifter-training method for narrow-band VC can shorten the tap length to 1/16 without degrading the converted-speech quality, and 2) the proposed sub-band modeling method for full-band VC can improve the converted-speech quality while reducing the computational cost, and 3) our real-time, online, full-band VC system can convert 48 kHz-sampled speech in real time attaining the converted speech with a 3.6 out of 5.0 mean opinion score of naturalness.

This paper proposes a change detection method for buildings based on convolutional neural networks. The proposed method detects building changes from pairs of optical aerial images and past map information concerning buildings. Using high-resolution image pair and past map information seamlessly, the proposed method can capture the building areas more precisely compared to a conventional method. Our experimental results show that the proposed method outperforms the conventional change detection method that uses optical aerial images to detect building changes.

This letter presents an efficient technique to reduce the computational complexity involved in training binary convolutional neural networks (BCNN). The BCNN training shall be conducted focusing on the optimization of the sign of each weight element rather than the exact value itself in convention; in which, the sign of an element is not likely to be flipped anymore after it has been updated to have such a large magnitude to be clipped out. The proposed technique does not update such elements that have been clipped out and eliminates the computations involved in their optimization accordingly. The complexity reduction by the proposed technique is as high as 25.52% in training the BCNN model for the CIFAR-10 classification task, while the accuracy is maintained without severe degradation.

This study is focused on the automated detection of a complex system operator's condition. For example, in this study a person's reaction while listening to music (or not listening at all) was determined. For this purpose various well-known data mining tools as well as ones developed by authors were used. To be more specific, the following techniques were developed and applied for the mentioned problems: artificial neural networks and fuzzy rule-based classifiers. The neural networks were generated by two modifications of the Differential Evolution algorithm based on the NSGA and MOEA/D schemes, proposed for solving multi-objective optimization problems. Fuzzy logic systems were generated by the population-based algorithm called Co-Operation of Biology Related Algorithms or COBRA. However, firstly each person's state was monitored. Thus, databases for problems described in this study were obtained by using non-contact Doppler sensors. Experimental results demonstrated that automatically generated neural networks and fuzzy rule-based classifiers can properly determine the human condition and reaction. Besides, proposed approaches outperformed alternative data mining tools. However, it was established that fuzzy rule-based classifiers are more accurate and interpretable than neural networks. Thus, they can be used for solving more complex problems related to the automated detection of an operator's condition.

This work presents a DNN accelerator architecture specifically designed for performing efficient inference on compressed and sparse DNN models. Leveraging the data sparsity, a runtime processing scheme is proposed to deal with the encoded weights and activations directly in the compressed domain without decompressing. Furthermore, a new data flow is proposed to facilitate the reusage of input activations across the fully-connected (FC) layers. The proposed design is implemented and verified using the Xilinx Virtex-7 FPGA. Experimental results show it achieves 1.99×, 1.95× faster and 20.38×, 3.04× more energy efficient than CPU and mGPU platforms, respectively, running AlexNet.

Recently, new optimization machines based on non-silicon physical systems, such as quantum annealing machines, have been developed, and their commercialization has been started. These machines solve the problems by searching the state of the Ising spins, which minimizes the Ising Hamiltonian. Such a property of minimization of the Ising Hamiltonian can be applied to various combinatorial optimization problems. In this paper, we introduce the coherent Ising machine (CIM), which can solve the problems in a milli-second order, and has higher performance than the quantum annealing machines especially on the problems with dense mutual connections in the corresponding Ising model. We explain how a target problem can be implemented on the CIM, based on the optimization scheme using the mutually connected neural networks. We apply the CIM to traveling salesman problems as an example benchmark, and show experimental results of the real machine of the CIM. We also apply the CIM to several combinatorial optimization problems in wireless communication systems, such as channel assignment problems. The CIM's ultra-fast optimization may enable a real-time optimization of various communication systems even in a dynamic communication environment.

Knowledge distillation approaches can transfer information from a large network (teacher network) to a small network (student network) to compress and accelerate deep neural networks. This paper proposes a novel knowledge distillation approach called multi-knowledge distillation (MKD). MKD consists of two stages. In the first stage, it employs autoencoders to learn compact and precise representations of the feature maps (FM) from the teacher network and the student network, these representations can be treated as the essential of the FM, i.e., EFM. In the second stage, MKD utilizes multiple kinds of knowledge, i.e., the magnitude of individual sample's EFM and the similarity relationships among several samples' EFM to enhance the generalization ability of the student network. Compared with previous approaches that employ FM or the handcrafted features from FM, the EFM learned from autoencoders can be transferred more efficiently and reliably. Furthermore, the rich information provided by the multiple kinds of knowledge guarantees the student network to mimic the teacher network as closely as possible. Experimental results also show that MKD is superior to the-state-of-arts.

A model extraction attack is a security issue in deep neural networks (DNNs). Information on a trained DNN model is an attractive target for an adversary not only in terms of intellectual property but also of security. Thus, an adversary tries to reveal the sensitive information contained in the trained DNN model from machine-learning services. Previous studies on model extraction attacks assumed that the victim provides a machine-learning cloud service and the adversary accesses the service through formal queries. However, when a DNN model is implemented on an edge device, adversaries can physically access the device and try to reveal the sensitive information contained in the implemented DNN model. We call these physical model extraction attacks model reverse-engineering (MRE) attacks to distinguish them from attacks on cloud services. Power side-channel analyses are often used in MRE attacks to reveal the internal operation from power consumption or electromagnetic leakage. Previous studies, including ours, evaluated MRE attacks against several types of DNN processors with power side-channel analyses. In this paper, information leakage from a systolic array which is used for the matrix multiplication unit in the DNN processors is evaluated. We utilized correlation power analysis (CPA) for the MRE attack and reveal weight parameters of a DNN model from the systolic array. Two types of the systolic array were implemented on field-programmable gate array (FPGA) to demonstrate that CPA reveals weight parameters from those systolic arrays. In addition, we applied an extended analysis approach called “chain CPA” for robust CPA analysis against the systolic arrays. Our experimental results indicate that an adversary can reveal trained model parameters from a DNN accelerator even if the DNN model parameters in the off-chip bus are protected with data encryption. Countermeasures against side-channel leaks will be important for implementing a DNN accelerator on a FPGA or application-specific integrated circuit (ASIC).

Recurrent neural networks (RNNs) have been proven effective for sequence-based tasks thanks to their capability to process temporal information. In real-world systems, deep RNNs are more widely used to solve complicated tasks such as large-scale speech recognition and machine translation. However, the implementation of deep RNNs on traditional hardware platforms is inefficient due to long-range temporal dependence and irregular computation patterns within RNNs. This inefficiency manifests itself in the proportional increase in the latency of RNN inference with respect to the number of layers of deep RNNs on CPUs and GPUs. Previous work has focused mostly on optimizing and accelerating individual RNN cells. To make deep RNN inference fast and efficient, we propose an accelerator based on a multi-FPGA platform called Flow-in-Cloud (FiC). In this work, we show that the parallelism provided by the multi-FPGA system can be taken advantage of to scale up the inference of deep RNNs, by partitioning a large model onto several FPGAs, so that the latency stays close to constant with respect to increasing number of RNN layers. For single-layer and four-layer RNNs, our implementation achieves 31x and 61x speedup compared with an Intel CPU.

Deep neural network (DNN)-based speech synthesis became popular in recent years and is expected to soon be widely used in embedded devices and environments with limited computing resources. The key intention of these systems in poor computing environments is to reduce the computational cost of generating speech parameter sequences while maintaining voice quality. However, reducing computational costs is challenging for two primary conventional DNN-based methods used for modeling speech parameter sequences. In feed-forward neural networks (FFNNs) with maximum likelihood parameter generation (MLPG), the MLPG reconstructs the temporal structure of the speech parameter sequences ignored by FFNNs but requires additional computational cost according to the sequence length. In recurrent neural networks, the recursive structure allows for the generation of speech parameter sequences while considering temporal structures without the MLPG, but increases the computational cost compared to FFNNs. We propose a new approach for DNNs to acquire parameters captured from the temporal structure by backpropagating the errors of multiple attributes of the temporal sequence via the loss function. This method enables FFNNs to generate speech parameter sequences by considering their temporal structure without the MLPG. We generated the fundamental frequency sequence and the mel-cepstrum sequence with our proposed method and conventional methods, and then synthesized and subjectively evaluated the speeches from these sequences. The proposed method enables even FFNNs that work on a frame-by-frame basis to generate speech parameter sequences by considering the temporal structure and to generate sequences perceptually superior to those from the conventional methods.

Machine learning models inherently memorize significant amounts of information, and thus hiding not only prediction processes but also trained models, i.e., model obliviousness, is desirable in the cloud setting. Several works achieved model obliviousness with the MNIST dataset, but datasets that include complicated samples, e.g., CIFAR-10 and CIFAR-100, are also used in actual applications, such as face recognition. Secret sharing-based secure prediction for CIFAR-10 is difficult to achieve. When a deep layer architecture such as CNN is used, the calculation error when performing secret calculation becomes large and the accuracy deteriorates. In addition, if detailed calculations are performed to improve accuracy, a large amount of calculation is required. Therefore, even if the conventional method is applied to CNN as it is, good results as described in the paper cannot be obtained. In this paper, we propose two approaches to solve this problem. Firstly, we propose a new protocol named Batch-normalizedActivation that combines BatchNormalization and Activation. Since BatchNormalization includes real number operations, when performing secret calculation, parameters must be converted into integers, which causes a calculation error and decrease accuracy. By using our protocol, calculation errors can be eliminated, and accuracy degradation can be eliminated. Further, the processing is simplified, and the amount of calculation is reduced. Secondly, we explore a secret computation friendly and high accuracy architecture. Related works use a low-accuracy, simple architecture, but in reality, a high accuracy architecture should be used. Therefore, we also explored a high accuracy architecture for the CIFAR10 dataset. Our proposed protocol can compute prediction of CIFAR-10 within 15.05 seconds with 87.36% accuracy while providing model obliviousness.

In this letter, a piecewise linear (PWL) sigmoid function approximation based on the statistical distribution probability of the neurons' values in each layer is proposed to improve the network recognition accuracy with only addition circuit. The sigmoid function is first divided into three fixed regions, and then according to the neurons' values distribution probability, the curve in each region is segmented into sub-regions to reduce the approximation error and improve the recognition accuracy. Experiments performed on Xilinx's FPGA-XC7A200T for MNIST and CIFAR-10 datasets show that the proposed method achieves 97.45% recognition accuracy in DNN, 98.42% in CNN on MNIST and 72.22% on CIFAR-10, up to 0.84%, 0.57% and 2.01% higher than other approximation methods with only addition circuit.

This paper presents a method for many-to-one voice conversion using phonetic posteriorgrams (PPGs) based on an adversarial training of deep neural networks (DNNs). A conventional method for many-to-one VC can learn a mapping function from input acoustic features to target acoustic features through separately trained DNN-based speech recognition and synthesis models. However, 1) the differences among speakers observed in PPGs and 2) an over-smoothing effect of generated acoustic features degrade the converted speech quality. Our method performs a domain-adversarial training of the recognition model for reducing the PPG differences. In addition, it incorporates a generative adversarial network into the training of the synthesis model for alleviating the over-smoothing effect. Unlike the conventional method, ours jointly trains the recognition and synthesis models so that they are optimized for many-to-one VC. Experimental evaluation demonstrates that the proposed method significantly improves the converted speech quality compared with conventional VC methods.

Some non-acoustic modalities have the ability to reveal certain speech attributes that can be used for synthesizing speech signals without acoustic signals. This study validated the use of ultrasonic Doppler frequency shifts caused by facial movements to implement a silent speech interface system. A 40kHz ultrasonic beam is incident to a speaker's mouth region. The features derived from the demodulated received signals were used to estimate the speech parameters. A nonlinear regression approach was employed in this estimation where the relationship between ultrasonic features and corresponding speech is represented by deep neural networks (DNN). In this study, we investigated the discrepancies between the ultrasonic signals of audible and silent speech to validate the possibility for totally silent communication. Since reference speech signals are not available in silently mouthed ultrasonic signals, a nearest-neighbor search and alignment method was proposed, wherein alignment was achieved by determining the optimal pair of ultrasonic and audible features in the sense of a minimum mean square error criterion. The experimental results showed that the performance of the ultrasonic Doppler-based method was superior to that of EMG-based speech estimation, and was comparable to an image-based method.

This paper proposes a method for heatmapping people who are involved in a group activity. Such people grouping is useful for understanding group activities. In prior work, people grouping is performed based on simple inflexible rules and schemes (e.g., based on proximity among people and with models representing only a constant number of people). In addition, several previous grouping methods require the results of action recognition for individual people, which may include erroneous results. On the other hand, our proposed heatmapping method can group any number of people who dynamically change their deployment. Our method can work independently of individual action recognition. A deep network for our proposed method consists of two input streams (i.e., RGB and human bounding-box images). This network outputs a heatmap representing pixelwise confidence values of the people grouping. Extensive exploration of appropriate parameters was conducted in order to optimize the input bounding-box images. As a result, we demonstrate the effectiveness of the proposed method for heatmapping people involved in group activities.

The increase in computation cost and storage of convolutional neural networks (CNNs) severely hinders their applications on limited-resources devices in recent years. As a result, there is impending necessity to accelerate the networks by certain methods. In this paper, we propose a loss-driven method to prune redundant channels of CNNs. It identifies unimportant channels by using Taylor expansion technique regarding to scaling and shifting factors, and prunes those channels by fixed percentile threshold. By doing so, we obtain a compact network with less parameters and FLOPs consumption. In experimental section, we evaluate the proposed method in CIFAR datasets with several popular networks, including VGG-19, DenseNet-40 and ResNet-164, and experimental results demonstrate the proposed method is able to prune over 70% channels and parameters with no performance loss. Moreover, iterative pruning could be used to obtain more compact network.

The neural network has been one of the most useful techniques in the area of speech recognition, language translation and image analysis in recent years. Long Short-Term Memory (LSTM), a popular type of recurrent neural networks (RNNs), has been widely implemented on CPUs and GPUs. However, those software implementations offer a poor parallelism while the existing hardware implementations lack in configurability. In order to make up for this gap, a highly configurable 7.62 GOP/s hardware implementation for LSTM is proposed in this paper. To achieve the goal, the work flow is carefully arranged to make the design compact and high-throughput; the structure is carefully organized to make the design configurable; the data buffering and compression strategy is carefully chosen to lower the bandwidth without increasing the complexity of structure; the data type, logistic sigmoid (σ) function and hyperbolic tangent (tanh) function is carefully optimized to balance the hardware cost and accuracy. This work achieves a performance of 7.62 GOP/s @ 238 MHz on XCZU6EG FPGA, which takes only 3K look-up table (LUT). Compared with the implementation on Intel Xeon E5-2620 CPU @ 2.10GHz, this work achieves about 90× speedup for small networks and 25× speed-up for large ones. The consumption of resources is also much less than that of the state-of-the-art works.