Developing an explainable Deep Neural Network for stress detection using biosignals and human-engineered features

With the ubiquity of handwriting in everyday tasks, it is surprising that existing avenues for handwriting authentication remain largely out of reach for the average individual or organization. Current solutions often rely on expensive or specialized equipment, and most existing research focuses on signatures rather than freeform handwriting. This limits the applicability of such technology to a narrow range of scenarios. In this paper, we argue that wearable devices might make handwriting authentication scalable and affordable. We design and evaluate two wearables-driven freeform handwriting authentication systems, one centered on a deep neural network and the other using human-engineered features. Our authentication systems are thoroughly tested across three writing experiments (involving 53 participants) that were carefully mapped to typical writing scenarios. We show the best performing configuration to attain an equal error rate of 5.51%, suggesting the potential of this modality for use in a multi-modal authentication system. To evaluate how our authentication systems perform against attacks by determined attackers, we developed and evaluated two impostor attacks that correspond to highly likely attack vectors. We then show that certain authentication system configurations are resistant to the attack. This paper represents an important step toward consumer ready wearables-driven freeform handwriting authentication.

Obstructive sleep apnea (OSA) is a sleep disorder that causes partial or complete cessation of breathing during an individual's sleep. Various methods have been proposed to automatically detect OSA events, but little work has focused on predicting such events in advance, which is useful for the development of devices that regulate breathing during a patient's sleep. We propose four methods for sleep apnea prediction based on convolutional and long short-term memory neural networks (1D-CNN, ConvLSTM, 1D-CNN-LSTM and 2D-CNN-LSTM), which use raw data from three respiratory signals (nasal flow, abdominal and thoracic) sampled at 32Hz, without any human-engineered features. We predict OSA (apnea or hypopnea) and normal breathing events 30 seconds ahead using the prior 90 seconds' data. Our results on a dataset containing over 46,000 examples from 1,507 subjects show that all four models achieved promising accuracy ( 81%). The 1D-CNN-LSTM and 2D-CNN-LSTM were the best two performing models with accuracy, sensitivity and specificity over 83%, 81% and 85% respectively. These results show that OSA events can be accurately predicted in advance based on respiratory signals, opening up opportunities for the development of devices to preemptively regulate the airflow to sleepers to avoid these events. Furthermore, we demonstrate good prediction performance even when respiratory signals are downsampled by a factor of 32, to 1Hz, for which our proposed 1D-CNN-LSTM achieved 82.94% accuracy, 81.25% sensitivity and 84.63% specificity. This robustness to low sampling frequencies allows our algorithms to be implemented in devices with low storage capacity, making them suitable for at-home environments.

The current Synthetic Aperture Radar (SAR) target detection methods based on Convolutional Neural Network (CNN) rely on a large amount of slice-level labeled train samples. However, it takes a lot of labor and material resources to label the SAR images at slice-level. Compared to label samples at slice-level, it is easier to label them at image-level. The image-level label indicates whether the image contains the target of interest or not. In this paper, a semi-supervised SAR image target detection method based on CNN is proposed by using a small number of slice-level labeled samples and a large number of image-level labeled samples. The target detection network of this method consists of region proposal network and detection network. Firstly, the target detection network is trained using the slice-level labeled samples. After training convergence, the output slices constitute the candidate region set. Then, the image-level labeled clutter samples are input into the network and then the negative slices of the output are added to the candidate region set. Next, the image-level labeled target samples are input into the network as well. After selecting the positive and negative slices in the output of the network, they are added to the candidate region set. Finally, the detection network is trained using the updated candidate region set. The processes of updating candidate region set and training detection network alternate until convergence. The experimental results based on the measured data demonstrate that the performance of the proposed method is similar to the fully supervised training method using a much larger set of slice-level samples.

In Integrated Circuit (IC) manufacturing, defect detection is a necessary task. The small size and high density of IC, and the complex characteristics of various defects, which burdens the defect detection system. It is difficult for the existing detection methods to accurately detect various types of defects while ensuring the detection speed. We implement a deep convolutional neural network in IC defect detection, based on YOLOv5, we add a prediction head to detect objects at different scales. In addition, we also integrate the Squeeze-and-Excitation layer (SELayer) to help the detection network find attention region on scenarios with dense objects, extract key features, enhance the network’s ability to detect difficult-to-detect samples, and generally improve the accuracy of the detection network. Extensive experiments on the IC defect dataset made by us show that the AP result of YOLOv5 with integrated attention module in IC defect detection is 95.4%, which is 0.5% better than the network without SELayer. Compared with other detection networks, it has obvious accuracy advantage, which is encouraging and promising competitive.

Background. The disadvantage of spectrally efficient signals with frequency multiplexing is the occurrence of intersymbol interference, which is further aggravated when these signals propagate in frequency selective channels.
 Aim. The possibility and effectiveness of using neural network approaches for channel equalization and signal detection in communication systems using SEFDM signals has been assessed.
 Methods. A receiver structure for SEFDM systems based on a deep complex-valued convolutional neural network is proposed, which allows recovering bits from the temporal representation of the signal without using the fractional Fourier transform and inverting the cross-correlation matrix between frequency subcarriers. A two-stage network training scheme has been developed. Based on simulation modeling, a comparative analysis of the noise immunity of SEFDM systems was carried out both in a channel with white Gaussian noise and in a channel with Rayleigh fading, using classical and neural network receivers.
 Results. It is shown that there is no loss of noise immunity in channels with additive white Gaussian noise and an increase in noise immunity of the system up to 2 dB in the channel specified by the extended automotive model (3GPP-EVA).
 Conclusion. The effectiveness of using deep neural complex-valued convolutional networks as receivers for spectrally efficient communication systems, as well as their advantage over classical ones, is shown.
 Keywords – SEFDM; deep complex-valued convolutional neural network; turbocoding.

Massive multiple-input multiple-output (MIMO) is a key technology for emerging next-generation wireless systems. Utilizing large antenna arrays at base-stations, massive MIMO enables substantial spatial multiplexing gains by simultaneously serving a large number of users. However, the complexity in massive MIMO signal processing (e.g., data detection) increases rapidly with the number of users, making conventional hand-engineered algorithms less computationally efficient. Lowcomplexity massive MIMO detection algorithms, especially those inspired or aided by deep learning, have emerged as a promising solution. While there exist many MIMO detection algorithms, the aim of this magazine paper is to provide insight into how to leverage deep neural networks (DNN) for massive MIMO detection. We review recent developments in DNN-based MIMO detection that incorporate the domain knowledge of established MIMO detection algorithms with the learning capability of DNNs. We then present a comparison of the key numerical performance metrics of these works. We conclude by describing future research areas and applications of DNNs in massive MIMO receivers.

Hierarchical neural networks have been shown to be effective in learning representative image features and recognizing object classes. However, most existing networks combine the low/middle level cues for classification without accounting for any spatial structures. For applications such as understanding a scene, how the visual cues are spatially distributed in an image becomes essential for successful analysis. This paper extends the framework of deep neural networks by accounting for the structural cues in the visual signals. In particular, two kinds of neural networks have been proposed. First, we develop a multitask deep convolutional network, which simultaneously detects the presence of the target and the geometric attributes (location and orientation) of the target with respect to the region of interest. Second, a recurrent neuron layer is adopted for structured visual detection. The recurrent neurons can deal with the spatial distribution of visible cues belonging to an object whose shape or structure is difficult to explicitly define. Both the networks are demonstrated by the practical task of detecting lane boundaries in traffic scenes. The multitask convolutional neural network provides auxiliary geometric information to help the subsequent modeling of the given lane structures. The recurrent neural network automatically detects lane boundaries, including those areas containing no marks, without any explicit prior knowledge or secondary modeling.

This paper proposes heterogeneous modular deep neural network (DNN) to address a complex problem of detection of diabetic retinopathy and simultaneously the five types of abnormalities. The modular approach gives the advantage to extract class specific features for the classifier, which helps to outperform the classical convolutional neural networks. Moreover, the heterogeneous nature of modular DNN provides the economy in the overall architecture and also enables to extract region specific features which further contribute to higher accuracy of detection. Extensive simulation studies were performed using benchmark dataset DIARETDB0 and results show that the proposed approach performs better or equivalently good than the other standard approaches.

Real-time smoke detection is of great significance for early warning of fire, which can avoid the serious loss caused by fire. Detecting smoke in actual scenes is still a challenging task due to large variance of smoke color, texture, and shapes. Moreover, the smoke detection in the actual scene is faced with the difficulties in data collection and insufficient smoke datasets, and the smoke morphology is susceptible to environmental influences. To improve the performance of smoke detection and solve the problem of too few datasets in real scenes, this paper proposes a model that combines a deep convolutional generative adversarial network and a convolutional neural network (DCG-CNN) to extract smoke features and detection. The vibe algorithm was used to collect smoke and nonsmoke images in the dynamic scene and deep convolutional generative adversarial network (DCGAN) used these images to generate images that are as realistic as possible. Besides, we designed an improved convolutional neural network (CNN) model for extracting smoke features and smoke detection. The experimental results show that the method has a good detection performance on the smoke generated in the actual scenes and effectively reduces the false alarm rate.

The low frequency ultra-wideband synthetic aperture radar (LF UWB BSAR) has a perfect capability of penetrating the vegetation and obtaining the high-resolution SAR image, which can be used to detect the foliage-concealed target under the jungles. However, the LF UWB SAR image usually includes some strong scattering points of the non-targets, such as the thick tree trunks. A change detection method of the foliage-concealed targets based on the deep neural network is presented in this paper for the LF UWB SAR images. Using the method of the data augmentation and random cutting, the LF UWB SAR image dataset is obtained from the CARABAS-II SAR dataset, which contains the foliage-concealed targets with the annotation. Besides, a deep neural network is designed for the prediction and detection of the foliage-concealed target in the LF UWB SAR images. The input is the previously generated LF UWB SAR image dataset, and the output is 1 or 0 which is used as the target judgment. Finally, the deep neural network is trained and tested, and the accuracy of the deep neural network for the change detection of the foliage-concealed targets in the LF UWB SAR images is obtained. Experimental results prove that the presented algorithm cannot only achieve the change detection of the foliage-concealed target in LF UWB SAR images, but also can keep the fast detection of the multi-targets but with a higher precision.

Blood transfusion is of great importance in the medical field, as blood donation centers are responsible for collecting and distributing blood and blood products. Artificial neural network is a data processing system getting ideas from the human brain and designs a data structure that acts like a neuron using programming science and by creating a network between these neurons and combining large amounts of data with smart algorithms and their rapid processing, the network is trained. In this study, data is extracted from the blood transfusion service center and the perceptron neural network, RBF neural network, Fisher’s discrimination ratio and genetic algorithm were examined, and finally the highest possible accuracy from the neural network was achieved.

Sentiment analysis and sarcasm detection are specialized areas in the field of information retrieval and natural language processing. Sentiment classification is closely correlated with sarcasm detection, where people usually adopt sarcasm to highlight their negative feeling. This paper proposes a novel multi-task deep neural networks for joint sarcasm detection and sentiment analysis (MT_SS). MT_SS train both tasks jointly using bidirectional gated recurrent unit with attention network module to obtain task-specific local feature representation while using convolutional neural networks to obtain global feature representation. The experiments on two datasets show that our proposed model outperforms the state-of-the-art approaches.

The diseases in plants pose a devastating impact on initiating safety in the production of food and they can lead to a reduction in the quantity and quality of agricultural products. In most cases, plant diseases lead to no grain harvest. Thus, an automatic diagnosis of plant disease is highly recommended for determining agricultural information. Several techniques are devised for plant disease detection wherein deep learning is preferred due to its effective performance. Novel deep learning is presented to spot disease from rice crop images. Here, the rice plant image undergoes pre-processing to remove noise and artifacts contained in the image. Then, the segmentation is performed with Segmentation Network (SegNet) to produce segments. The segments are further adapted for extracting statistical features, convolution neural network (CNN) features and texture features. These features are employed for plant disease detection wherein the deep recurrent neural network (Deep RNN) is utilized. The Deep RNN is trained with the proposed RideSpider Water Wave (RSW) algorithm. The proposed RSW is devised by integrating RWW in Spider monkey optimization. The proposed RWS-based Deep RNN provides superior performance with the highest accuracy of 90.5%, maximal sensitivity of 84.9% and maximal specificity of 95.2%.

PurposeWe present image classifiers based on Dense Convolutional Networks and transfer learning to classify chest X-ray images according to three labels: COVID-19, pneumonia, and normal.MethodsWe fine-tuned neural networks pretrained on ImageNet and applied a twice transfer learning approach, using NIH ChestX-ray14 dataset as an intermediate step. We also suggested a novelty called output neuron keeping, which changes the twice transfer learning technique. In order to clarify the modus operandi of the models, we used Layer-wise Relevance Propagation (LRP) to generate heatmaps.ResultsWe were able to reach test accuracy of 100% on our test dataset. Twice transfer learning and output neuron keeping showed promising results improving performances, mainly in the beginning of the training process. Although LRP revealed that words on the X-rays can influence the networks’ predictions, we discovered this had only a very small effect on accuracy.ConclusionAlthough clinical studies and larger datasets are still needed to further ensure good generalization, the state-of-the-art performances we achieved show that, with the help of artificial intelligence, chest X-rays can become a cheap and accurate auxiliary method for COVID-19 diagnosis. Heatmaps generated by LRP improve the interpretability of the deep neural networks and indicate an analytical path for future research on diagnosis. Twice transfer learning with output neuron keeping improved DNN performance.

Automated methods for detecting clinically significant (CS) prostate cancer (PCa) in multi-parameter magnetic resonance images (mp-MRI) are of high demand. Existing methods typically employ several separate steps, each of which is optimized individually without considering the error tolerance of other steps. As a result, they could either involve unnecessary computational cost or suffer from errors accumulated over steps. In this paper, we present an automated CS PCa detection system, where all steps are optimized jointly in an end-to-end trainable deep neural network. The proposed neural network consists of concatenated subnets: 1) a novel tissue deformation network (TDN) for automated prostate detection and multimodal registration and 2) a dual-path convolutional neural network (CNN) for CS PCa detection. Three types of loss functions, i.e., classification loss, inconsistency loss, and overlap loss, are employed for optimizing all parameters of the proposed TDN and CNN. In the training phase, the two nets mutually affect each other and effectively guide registration and extraction of representative CS PCa-relevant features to achieve results with sufficient accuracy. The entire network is trained in a weakly supervised manner by providing only image-level annotations (i.e., presence/absence of PCa) without exact priors of lesions' locations. Compared with most existing systems which require supervised labels, e.g., manual delineation of PCa lesions, it is much more convenient for clinical usage. Comprehensive evaluation based on fivefold cross validation using 360 patient data demonstrates that our system achieves a high accuracy for CS PCa detection, i.e., a sensitivity of 0.6374 and 0.8978 at 0.1 and 1 false positives per normal/benign patient.