Search

In recent times, the use of chest X-ray (CXR) images in telemedicine has increased exponentially. However, the transfer of these images on an unsecured channel poses a serious threat to their authenticity and copyright protection. To address this problem, a chest X-ray image watermarking scheme (CXRmark) is presented in this paper using an online sequential reduced kernel extreme learning machine (OS-RKELM). U-Net is used to segment the lung area into the region of non-interest (RONI) and region of interest (ROI). An OS-RKELM is trained using the approximation coefficients (obtained by transforming training images using discrete wavelet transform) and the corresponding quantized coefficients. Subsequently, it is used to modulate the approximation coefficients. In order to avoid inadvertent deterioration in the ROI and to achieve high robustness, the proposed watermarking scheme uses different embedding strengths for ROI and RONI. Further, to enhance the security of the watermark, it has been encoded using the logistic chaotic map. To evaluate the performance of the proposed scheme, we experimented with a set of 461 CXR images from a publicly available repository: COVID-19 Image Data Collection. It has been demonstrated that CXRmark results in high perceptual quality of the signed CXRs and is robust to common interference. Further, the reversibility of the proposed CXRmark scheme is demonstrated by recovering the undistorted CXR image from the signed CXR. The comparison of the proposed scheme with state-of-the-art schemes shows that CXRmark outperforms its competitors in terms of imperceptibility and robustness. [ABSTRACT FROM AUTHOR]

Machine learning-based direction-of-arrival (DOA) estimation methods can have a good predictive ability even in complex scenarios. However, their estimation performance in the face of unknown data is poor because they rely on the generalization ability of the algorithm itself. Therefore, this paper proposes a DOA estimation method based on the Light Gradient Boosting Machine (LightGBM) algorithm. Using the histogram algorithm, gradient-based one-sided sampling and exclusive feature bundling measures, the LightGBM algorithm can reduce the time to find the best segmentation point, reduce the amount of data and the number of features in the dataset, thus reducing the model training time, and achieve high prediction accuracy. Applying the LightGBM algorithm to the DOA estimation problem and using a large dataset for training can improve the estimation accuracy while reducing the training cost. Simulation and real experimental results verify the effectiveness of the method. [ABSTRACT FROM AUTHOR]

The ever-increasing use of artificial intelligence (AI) to optimize embedding a watermark or find the optimal location in the host image for inserting a watermark might solve some problems in this field. However, the main problem, which is finding the optimum trade-off point among several watermarking criteria, has still not been investigated by researchers in this field, especially for speech signals. This paper aims to find the best trade-off among the watermarking requirements such as capacity, inaudibility, and robustness by applying an AI model. Moreover, a novel watermarking technique is proposed by modification of the probability density function (PDF) of the linear predictive (LP) residual and wavelet detail coefficient. For this method, a mathematical model is developed based on applying higher-order statistics for embedding and extracting the watermark. Sinh-arcsinh is used to shape the skewness and kurtosis of normal distribution for the LP residual or the wavelet high-frequency sub-bands, respectively, based on the watermark bits. Experimental results will show that although LP residual is not robust and it shows random behavior for modeling its PDF, the wavelet high-frequency band is quite robust and it can model the PDF of the wavelet. Moreover, it is demonstrated that AI has the capability to compromise among the watermarking criteria. Conclusions are drawn based on theoretical (maximum likelihood) and AI (machine learning) approaches, which confirm the effectiveness of the proposed model. Finally, in conclusion, several potential areas are discussed for further exploration. [ABSTRACT FROM AUTHOR]

The path planning strategies of multi-objective tasks for autonomous agents are applied more and more widely in the areas of resource exploration, factory transportation, post-disaster relief, etc. These tasks typically require access to all target points within a limited time and can be formulated by signal temporal logic (STL) formulas with multiple logic AND. This paper presents a dual-layer network deep reinforcement learning algorithm to address the decision problem of the multi-objective STL tasks. First, the multi-objective optimization problem is represented by iterable UNTIL operators, and each sub-task is encoded into an one-hot vector. In the control layer, one-hot vector is updated according to the ongoing sub-task. Based on the coupling input of agent states and one-hot vector, the neural network controller is trained to complete the sub-tasks in the given order strictly. In the order layer, robust-based weighted entropy is assigned according to the agent trajectory robust value of each sub-task. By maximizing the entropy, the order priority of access to the target points is updated periodicity and is passed into the control layer. Simulation and comparison with the traditional method show that our proposed dual-layer network algorithm has the reliability for strictly completing sub-tasks of multi-objective task reformed by UNTIL operators and the effectiveness of the access order planning for multi-objective tasks with improved real-time performance. [ABSTRACT FROM AUTHOR]

In this paper, a novel method is introduced to diagnose the cardiac arrhythmias. Automatic diagnosis of cardiac arrhythmias is crucial for successful treatment of heart diseases, and nowadays, machine learning is used to facilitate this purpose. For accurate classification of arrhythmias, it is vital to extract some appropriate features which distinguish different classes. In this research, a deep convolutional neural network is used to extract the features. Since the heart rates of various patients are very different, arrhythmias of a specific class will be placed in different locations of feature space. To reduce the intra-class variations, each patient's heart rate is mapped with a dedicated function to increase its similarity to the heart rate of a training patient. The proposed patient-specific mapping decreases the intra-class variation and significantly increases the classification accuracy of cardiac arrhythmias. The calculated accuracy is about 91.08%, which shows the better performance compared to other similar works. [ABSTRACT FROM AUTHOR]

This work investigates the significance of Hilbert domain characterization of wavelet packets in classifying different emotion of speech signal. The goal of this paper is to create a new emotional speech database and introduce a new feature extraction approach that can recognize various emotions. The proposed feature, wavelet cepstral coefficients (WCC) are based on Hilbert spectrum analysis of the wavelet packet of the speech signal. The speaker-independent machine learning models are developed using multiclass support vector machine (SVM) and k-nearest neighbourhood (KNN) classifier. The approach is tested with newly developed Telugu Indian database and the EMOVO (Italian emotional speech) database. Our proposed wavelet features achieve a peak accuracy of 73.5%, further boosted by NCA feature selection by 3–5%, resulting in an improved unweighted average recall (UAR) of 78% for database 1 and 87.50% for database 2, employing optimal wavelet features in conjunction with SVM classification. The proposed features outperformed the baseline Mel-frequency cepstral coefficients (MFCC) feature. The performance of newly formulated features is better than other existing methodologies tested with different language databases. [ABSTRACT FROM AUTHOR]

Several works report on overcoming the challenge of complex general-purpose signal processing algorithms that result in quick utilization of chip area resources (like FPGA in implementing system on chip or network on chip). However, with the advent of machine learning and deep learning algorithms, reducing this complexity and immense use of resources is a challenging task that can improve the system's performance in real time. One of the approaches is algorithm optimization, which consumes fewer resources and produces an equivalent performance. This approach may likely be accepted in less-sensitive applications, like voice or video processing. Recently, an optimized algorithm of correlation-less Wiener–Hopf-based adaptive channel equalizer is reported, in which the overall complexity of the system is reduced by employing a more compact way of weight optimization. This paper is the continuation of that work that reports the modification in the Steepest-Descent-based adaptive channel equalizer algorithm by keeping the autocorrelation matrix out of the algorithm. The proposed design is simulated and tested at equivalent spectral performance and produces comparable performance with fewer chip resources when translated on FPGA. The proposed design supports that algorithm-level optimization may be accepted extensively for the optimal hardware-based design of DSP systems. [ABSTRACT FROM AUTHOR]

Traditional von Neumann architecture-based processors become inefficient in terms of energy and throughput as they involved separate processing and memory units, also known as memory walls. The memory wall problem is further exacerbated when massive parallelism and frequent data movement are required between processing and memory units for real-time implementation of artificial neural networks (ANNs) that enable many intelligent applications. One of the most promising approach to address this problem is to perform computations inside the memory core that enhances the memory bandwidth and energy efficiency. This paper presents an in-memory computing architecture for ANN enabling artificial intelligence (AI) and machine learning (ML) applications. The proposed architecture utilizes a standard six-transistor (6T) static random access memory (SRAM) core for the implementation of a multilayered perceptron. Our novel on-chip training and inference in-memory architecture reduces energy cost and enhances throughput by simultaneously accessing the multiple rows of SRAM array per pre-charge cycle by eliminating frequent data access. The proposed architecture was trained and tested on the IRIS dataset and was observed to consume ≈ 22.46 × less energy/decision compared to earlier DIMA-based classifiers. [ABSTRACT FROM AUTHOR]

This paper presents a least-mean-square-based (LMS-based) background calibration algorithm with reinforcement learning optimization to calibrate the capacitor mismatch in successive approximation-register (SAR) analog-to-digital converters (ADCs). When calibrating capacitor mismatch, the convergence speed and stability of the conventional LMS-based algorithm cannot be balanced due to the fixed iteration step size. To improve the calibration effect, nine different iteration step sizes are proposed for calibration. An iteration step size selection strategy is obtained by training the Q-learning algorithm with the conventional LMS-based calibration iteration process. Leading by this strategy, the proposed calibration algorithm performs better convergence speed and stability than the conventional LMS-based calibration algorithm. A 12-bit sub-radix-2 redundant SAR ADC is used to test our proposed calibration algorithm for better evaluation. Behavioral simulations show that the proposed calibration algorithm converges at 0.065 M sampling points, and the standard deviation of the effective number of bits (ENOB) in the convergence state is only 0.034-bit. [ABSTRACT FROM AUTHOR]

Sound source localization is a technique that utilizes microphone arrays to detect the position of sound sources. It has a wide range of applications in areas such as smart homes, robot navigation, and conference recording. However, due to the complexity of the acoustic environment and the impact of noise interference, the accuracy of localization algorithms has always been a core concern in this field. Traditional time difference of arrival (TDOA) techniques struggle to achieve high precision and efficiency in localization results. To address this issue, this paper proposes a microphone array sound source localization method based on incremental broad learning (Enhance) algorithm. This method extracts shallow and deep features from audio signals and maps them to feature nodes and enhancement nodes in the broad learning system (BLS); a neural network model is constructed, which allows for fast adjustment of network structure and parameters. The model employs ridge regression to calculate connection weights and utilizes enhancement nodes to modify and optimize the feature nodes, thus achieving accurate prediction of sound source locations. The proposed method is experimentally validated using the NOIZEUS dataset and compared with the single-structure broad learning (One-shot) algorithm, back-propagation neural network (BP) algorithm, and recurrent neural network (RNN) algorithm. In the experiments, a microphone array consisting of four microphones is used, with a room size of 5 m × 4 m × 3 m. Different reverberation times (T60) and signal-to-noise ratios (SNRs) are employed to simulate various acoustic environments, and the performance of the four algorithms is evaluated in terms of outlier percentage and mean-squared error (MSE). The results demonstrate that under high reverberation (T60 = 700 ms) and low SNR (SNR = 0 dB) conditions, the proposed method achieves outlier percentage of only 0.308% and MSE of 0.92°. Compared to the other three algorithms, Enhance algorithm exhibits superior localization accuracy, noise robustness, and stability, thus holding significant research and practical value. [ABSTRACT FROM AUTHOR]

Multimedia data is crucial in the military, medical, forensics, social, etc., to transmit a large amount of data. Security of this sensitive information is the primary issue. This paper uses Latin square and machine learning techniques such as neural networks and genetic algorithm to design an image encryption algorithm. A new neural network-based pseudorandom number generator is proposed to generate a chaotic sequence for various applications. Encryption key images are designed using Latin squares in the finite field. Further, the Latin squares are XOR with the input matrix to get the encrypted images. The proposed algorithm is iterated a finite number of times to generate a cipher image population. Randomly two parents are chosen from the generated population, and row and column arrangements produce offspring. A genetic algorithm is the optimization technique used for the best encrypted image search. The pixel correlation value serves as a fitness function. Finally, the least correlated cipher image is obtained from the genetic algorithm applied to the parent and offspring of the population generated from the encryption algorithm. The simulation results from the proposed image encryption model surpass many communication channel attacks and perform better when compared to existing image security algorithms. [ABSTRACT FROM AUTHOR]

Bipolar disorder is a serious psychiatric disorder characterized by periodic episodes of manic and depressive symptomatology. Due to the high percentage of people suffering from severe bipolar and depressive disorders, the modelling, characterisation, classification and diagnostic analysis of these mental disorders are of vital importance in medical research. Electroencephalogram (EEG) records offer important information to enhance clinical diagnosis and are widely used in hospitals. For this reason, EEG records and patient data from the Virgen de la Luz Hospital were used in this work. In this paper, an extreme gradient boosting (XGB) machine learning (ML) method involving an EEG signal is proposed. Four supervised ML algorithms including a k-nearest neighbours (KNN), decision tree (DT), Gaussian Naïve Bayes (GNB) and support vector machine (SVM) were compared with the proposed XGB method. The performance of these methods was tested implementing a standard 10-fold cross-validation process. The results indicate that the XGB has the best prediction accuracy (94%), high precision ( > 0.94 ) and high recall ( > 0.94 ). The KNN, SVM, and DT approaches also present moderate prediction accuracy ( > 87 ), moderate recall ( > 0.87 ) and moderate precision ( > 0.87 ). The GNB algorithm shows relatively low classification performance. Based on these results for classification performance and prediction accuracy, the XGB is a solid candidate for a correct classification of patients with bipolar disorder. These findings suggest that XGB system trained with clinical data may serve as a new tool to assist in the diagnosis of patients with bipolar disorder. [ABSTRACT FROM AUTHOR]

At present, the deep learning super-resolution (SR) method has achieved excellent results, but it also faces problems such as large models, high computational cost, a large amounts of training data, and poor interpretability. However, traditional machine learning-based methods still have room for improvement in feature extraction and model structure. This paper constructs a gradient embedding cascade forest structure on the basis of random forest and proposes a limit gradient embedding cascaded forest SR (LGECFSR) model. In feature construction, we not only adopt the first-order gradient, the second-order gradient, and other features of the image but also fuse the information of the original LR image. In addition, image blocks of different sizes are used for training, which increases the model's generalization ability. Compared with the state-of-the-art machine learning-based methods, our method achieves the best performance and the second-best computational speed. In addition, compared with some deep learning-based methods, our model has a similar reconstruction effect and the best computational speed. In detail, for some reconstruction tasks, the Multi-Adds of LGECFSR is one-tenth to one-4000th of that of some current models. However, the SR performance of LGECFSR is the same or slightly better than that of some current classical algorithms. [ABSTRACT FROM AUTHOR]

Correct diagnosing analog circuit fault is beneficial to the circuit's health management, and its core challenge is extracting essential features from the circuit's output signals. Wavelet transform is a classical features extraction method whose performance relies on its wavelet basis function deeply. However, there are no satisfying rules to discover an optimal wavelet basis function for wavelet transform. In this paper, an improved wavelet transform with optimal wavelet basis function selection strategy is proposed. In the strategy, the optimal wavelet basis function is selected based on calculating the distance score and mean score of its features, and the features extracted by the optimal wavelet basis function are considered as the best features of signals. Subsequently, the features are split into training data and testing data randomly and evenly. By using the training data, a multiple kernel extreme learning machine (MKELM) based diagnosing model is initialized, and the parameters of MKELM are yielded by using particle swarm optimization algorithm. Finally, the MKELM is used to identify the faults of testing data for the purpose of verifying its performance. Fault diagnosis experiments of three circuits are performed to show the proposed optimal wavelet basis function selection strategy and MKELM's establishing process. Comparison experiments are performed to verify that the optimal wavelet basis function selection strategy is effective and MKELM is better than other classifiers in analog circuit fault diagnosis. [ABSTRACT FROM AUTHOR]

The complex kernel adaptive filter (CKAF) has been widely applied to the complex-valued nonlinear problem in signal processing and machine learning. However, most of the CKAF applications involve the complex kernel least mean square (CKLMS) algorithms, which work in a pure complex or complexified reproducing kernel Hilbert space (RKHS). In this paper, we propose the generalized complex kernel affine projection (GCKAP) algorithms in the widely linear complex-valued RKHS (WL-RKHS). The proposed algorithms have two main notable features. One is that they provide a complete solution for both circular and non-circular complex nonlinear problems and show many performance improvements over the CKAP algorithms. The other is that the GCKAP algorithms inherit the simplicity of the CKLMS algorithm while reducing its gradient noise and boosting its convergence. The second-order statistical characteristics of WL-RKHS have also been developed. An augmented Gram matrix consists of a standard Gram matrix and a pseudo-Gram matrix. This decomposition provides more underlying information when the real and imaginary parts of the signal are correlated and learning is independent. In addition, some online sparsification criteria are compared comprehensively in the GCKAP algorithms, including the novelty criterion, the coherence criterion, and the angle criterion. Finally, two nonlinear channel equalization experiments with non-circular complex inputs are presented to illustrate the performance improvements of the proposed algorithms. [ABSTRACT FROM AUTHOR]

In this article, a novel combined i-vector and an Extreme Learning Machine (ELM) is proposed for speaker identification. The ELM is chosen because it is fast to train and has a universal approximator property. Four combinations of features based on Mel Frequency Cepstral Coefficient and Power Normalized Cepstral Coefficient are used. Besides, seven fusion methods are exploited. The system is evaluated with three different databases, namely: the SITW 2006, NIST 2008, and the TIMIT database. This work employs the 2016 SITW database for the first time for speaker identification using the integration between the ELM and i-vector approach. From each database, 120 speakers with 1200 speech utterances are used (overall 360 speakers with 3600 speech utterances). Furthermore, comprehensive evaluations are exploited with a wide range of realistic background noise types (Stationary noise AWGN and Non-Stationary Noise types) with the handset effect. The proposed system is compared with the Gaussian Mixture Model-Universal Background Model (GMM-UBM) and other states of the art approaches. The results show that the i-vector method outperforms the GMM-UBM approach and other state- of-the-art methods under specific conditions, and that fusion techniques can be used to improve robustness to noise and handset effects. [ABSTRACT FROM AUTHOR]

In this work, we investigate epilepsy seizure detection using machine learning. In the literature, a machine learning model is usually trained to help automate the epileptic detection process, eliminating the need for human intervention. Typically, the dataset is split into training and test sets in a way to maximize the detection accuracy. This requires the training set to include enough EEG samples for every possible patient in order to improve the accuracy numbers during the prediction. However, this might not be easy or practical in real life. A new patient might not have a previous record in the training set, and hence, the prediction for this particular patient might not meet the expected accuracy. The main contribution in this work is to study the impact of the training and test datasets selection from practical point of view on the accuracy and efficacy of the CNN prediction. In this work, a CNN model, namely AlexNet, is trained to detect epileptic states, namely preictal, interictal and ictal, in subjects using electroencephalogram (EEG) signals. The dataset includes the three epileptic zones of subjects taken from three medical centers, collected by the Fragility Multi-Center Retrospective Study. Furthermore, we propose a framework to utilize a feature extraction technique that exploits the available multiple channels of EEG signals to minimize information loss. As part of the main contribution, three different approaches are proposed to split the EEG sample dataset into the training and test sets. Thus, the prediction performance is evaluated based on the prior knowledge extracted from the particular samples picked for the training set. The results show that the proposed framework achieves an overall accuracy of 94.44% when the training contained samples from each patient. The accuracy is reduced to 92.98% when the training set contained a subset of the patient pool. A binary classification is also performed with up to 98% accuracy for both scenarios. [ABSTRACT FROM AUTHOR]

Brain activities recorded using electroencephalography (EEG) device are mostly contaminated with eye blink (EB) artifact. This artifact leads to poor performance of brain–computer interface (BCI) systems. Hence, for the better performance of BCI systems, EB artifacts need to be removed from EEG signals without any loss of information. Of several methods that exists in the literature to remove EB artifacts, sparsity-based method is one among them and it proved to be good in removing EB artifacts. In the sparsity-based method, an over-complete dictionary is learned from the EEG data itself using K-SVD-based algorithm and is designed to model EB characteristics. In this work, two different greedy algorithms, namely orthogonal matching pursuit (OMP) and adaptive OMP (A-OMP), have been applied over K-SVD algorithm to check its performance on removing EB artifacts from EEG signals. To prove the efficiency of the greedy algorithms, the experiment is done with real EEG data. The results observed show that A-OMP is computationally more efficient and can accomplish successful sparse representation on EEG signals. Moreover, this sparsity-based algorithm can eliminate EB artifact accurately from the EEG signals. [ABSTRACT FROM AUTHOR]

Epilepsy is a neurological disorder directly linked with brain electrical activities, which causes sudden and recurrent seizures in the patient. Epilepsy seizures can be detected by using automatic detection systems by analyzing significant features extracted from EEG recordings. In this work, we aim to focus on adaptive mode decomposition methods, namely empirical mode decomposition (EMD), empirical wavelet transform (EWT) and variational mode decomposition (VMD) methods, that decompose EEG signals into different levels of resolution and enable extracting relevant nonlinear features for accurate detection of epilepsy seizures. We propose an intelligent epilepsy seizure detection system using a neural network (NN) classifier based on nonlinear features extracted from ECG signals using adaptive mode decomposition methods. In addition, we propose to use hybrid features selected using a wrapper-based feature selection method from nonlinear features extracted using different adaptive mode decomposition methods. The experimental results prove that the proposed system can detect epilepsy seizures up to an accuracy of 99%, the sensitivity of 98%, specificity of 99% and area under ROC (AUC) of 99% using NSC_ND dataset. We also conduct nonparametric statistical significance tests, Friedman test and Wilcoxon signed ranks post hoc test for demonstrating statistical differences between the obtained results and superior performance of the proposed system. This study enables researchers and practitioners to examine the proposed method and adaptive mode decomposition methods for detecting epilepsy seizures. [ABSTRACT FROM AUTHOR]

Acoustic Scene Classification (ASC) aims to categorize real-world audio into one of the predetermined classes that identifies the recording environment of the audio. State-of-the-art ASC algorithms have excellent performance in terms of accuracy due to the emergence of deep learning algorithms. In particular, Convolutional Neural Networks (CNN) have set a new benchmark in ASC due to their promising performance. Despite the emergence of new frameworks, the interest in ASC is growing progressively with a shift of focus from enhancing accuracy to reducing model complexity. In this work, we introduce the AtResNet, a residual atrous CNN for low complexity acoustic scene classification. The AtResNet utilizes dilated convolutions and residual connections to reduce the number of model parameters. To further enhance the performance of AtResNet, we introduce a multi-scale feature representation method called multi-scale mel spectrogram (ms2). To compute the ms2, we evaluate the mel spectrogram on the wavelet subbands of the signal. We assessed AtResNet with ms2 on three benchmark datasets in ASC. The results suggest that our method significantly outperformed the CNN-based techniques in addition to a baseline system based on log mel spectrum for signal representation. AtResNet offers a 28.73% reduction in the model parameters against a baseline CNN. Furthermore, the AtResNet has a model size of 81 KB with post-training quantization of network weights. It makes AtResNet suitable for deployment in context-aware devices. [ABSTRACT FROM AUTHOR]

Various authors have designed multirate filter banks for modeling a given process. The existing algorithms, however, suffer from some major drawbacks; for example, some of these filter banks are not signal adaptive; in most of the cases the, filter computations are not efficient; moreover, the existing signal-adapted filter banks is not always optimal. To address these limitations, we propose an optimal signal matched filter bank (SMFB) to maximize the coding gain. The objective is to achieve spectral majorization and complete decorrelation of the sub-band signals. The filter bank is designed here to perfectly reconstruct only the given signal, instead of every signal from the finite energy signal space, as is typically done. These objectives are achieved by whitening the outputs of the analysis filter bank, within as well as across the channels. With these orthogonal whitened signals as inputs, the synthesis filter bank is designed to reconstruct a delayed version of the given signal. A time and order recursive least squares algorithm has been developed for the SMFB, with a single realization of the signal. The recursions of the algorithm lead to a lattice structure. Since the filter coefficients are not directly available from the lattice, a fast algorithm is developed to estimate the same using the lattice parameters. Simulation results are presented to verify the efficacy of the proposed theory. We also present an automated real-time sleep apnea detection algorithm using SMFB which gives a 92.27 % accuracy, 88.60 % sensitivity, and 94.31 % specificity on a publicly available MIT PhysioNet Apnea-ECG dataset. [ABSTRACT FROM AUTHOR]

Through wearable technology, several chronic diseases are diagnosed by long-term monitoring of vital signs specifically ECG, EMG, EEG biosignals. Such prolonged monitoring and transmitting these multiple recordings may decline the battery power of wireless wearable device. This work aims at preserving the battery power of wireless wearables by jointly compressing ECG–EMG–EEG signals before sending to the receiver. This work proposes multimodal deep denoising convolutional autoencoder architecture for joint compression (encoding) and reconstruction (decoding) of ECG–EMG–EEG biosignals. In addition, the system may encounter new data stream in future with varying range of statistics in this real-time scenario; hence, it is required to remodel the system. But these wearables are memory constrained, so the model's learned optimized parameters should not increase in size when it is remodeled or updated. The incremental learning addresses this issue by reusing the previously learned weights as initial weight for retraining the model for new dataset and avoids random weight initialization thereby maintaining the space and time complexity. The experimental result shows that the proposed model achieves better compression efficiency of 99.8% with highest reconstruction Quality Score of 156, 254 & 149.4 for ECG, EMG & EEG signals, respectively, than state-of-the-art methods, and it is observed that the computation time is low for joint compression than compressing each signal individually. [ABSTRACT FROM AUTHOR]

This work presents a PVT robust machine learning-based Voltage-controlled Oscillator (VCO) with an enhanced linearity range. The machine learning algorithm with PVT robustness is implemented digitally. Different from conventional methods, the proposed scheme does not require the VCO to be in working mode every time one needs the prediction of frequency. The proposed scheme uses the frequency to digital converter (FDC) output data as an input learning vector and uses a prediction block to predict the future frequencies. An 11-stage voltage-controlled oscillator with a machine learning algorithm is implemented in SCL 180 nm CMOS technology. The measurement results show that the proposed architecture is robust against PVT variations with an enhanced linearity range. Without a machine learning algorithm, the VCO's control voltage linearity range is 0.28 V to 0.40 V that increases to the range from 0.1 to 1.8 V after applying the proposed machine learning algorithm. The maximum gain variation of 3.71% is observed at FF with respect to the TT corner after applying the proposed machine learning algorithm. [ABSTRACT FROM AUTHOR]

The most commonly used training target is masking-based approach which maps noisy speech to the time–frequency (T–F) unit and has a remarkable impact on the performance in the supervised learning algorithms. Traditional T–F masks like ideal ratio mask (IRM) demonstrate a strong performance but are limited to only the magnitude domain in enhancement. Though bounded IRM with phase constraint (BIRMP) includes phase difference but doesn't exploit channel correlation, the proposed ratio mask (pRM) considers channel correlation but is computed only in the magnitude domain. This work proposes a new mask, i.e., phase correlation ideal ratio mask (PCIRM), which includes both inter-channel correlation and phase difference between the noisy speech ( N S ), noise (N) and clean speech ( C S ). Considering these factors increases the percentage of C S and readily decreases the percentage of unwanted noise in the speech components and conversely for the noise components making the mask more precise. The experimental results are conducted under different SNR levels using TIMIT dataset and NOISEX-92 dataset and also compared with the existing state-of-the-art approaches. The results prove that the proposed mask has higher performance than BIRMP and pRM in terms of speech quality and intelligibility. [ABSTRACT FROM AUTHOR]

This work proposes a new artificial bandwidth extension (ABE) framework for enhancing the quality of narrowband speech signals. This enhancement process recovers missing high-frequency components of the signal. In this regard, a new bandwidth extension process based on H ∞ sampled-data control theory and machine learning models is proposed. In addition, a little non-ideality (aliasing) is allowed in the narrowband signal to get better reconstruction for the missing higher frequencies. The H ∞ sampled-data control theory works on a signal model, representing the already available wideband signal. Direct use of this theory is not possible in the bandwidth extension process as the signal models may not be the same for different phonemes of speech signals, even if uttered by the same speaker due to their non-stationary behavior. Hence, machine learning models are necessary. We have performed experiments with four types of narrowband features and two types of machine learning models approaches. The proposed method improves most of the measures when compared to the existing techniques, such as 12% minimum improvement in log spectral distance (LSD). [ABSTRACT FROM AUTHOR]

Deep nonnegative matrix factorization (DMF) is a particularly useful technique for learning data representation in low-dimensional space. To further obtain the complex hidden information and keep the geometrical structures of the high-dimensional data, we propose a novel deep matrix factorization model with the graph regularization (called DGsnMF). For solving the model with multi-variables, we design a forward–backward splitting scheme. After that, the convergence analysis is attached to the proposed algorithm and it is proved to converge to a critical point. Empirical experiments on benchmark datasets show that the proposed method is superior to the compared methods. [ABSTRACT FROM AUTHOR]