Non-autoregressive text-to-speech (TTS) has recently received a lot of atten-tion due to its reliability and fast reasoning. Despite its outstanding achieve-ment, non-autoregressive speech synthesis still faces some critical challenges. A major issue is that non-autoregressive methods necessitate an external toolkit to align the speech with the transcript, thus substantially complicat-ing the process of building the model. Besides, non-autoregressive methods suffer from the one-to-many mapping issue, where the same transcript may correspond to speech in numerous styles. This problem may harm the ex-pressiveness of the generated speech because the model tends to provide output with an average style. To address the above issues, this paper pro-poses a cooperative learning strategy for non-autoregressive speech synthe-sis. Specifically, the suggested method employs both an autoregressive and a non-autoregressive TTS model during the training procedure. The autore-gressive model is trained as a partner at each iteration, providing essential alignment information and also the prosody embedding of the speech to the non-autoregressive model. After receiving the above useful knowledge, the non-autoregressive model can be further trained without relying on external alignment tools. Meanwhile, the prosody embedding from the autoregressive model and the pitch information from the raw audio can be utilised together to alleviate the one-to-many mapping problem. Experimental results demon-strate that our approach can produce comparable speech to the most popular FastSpeech 2 model while drastically reducing the complexity of constructing a non-autoregressive TTS model.

Explainable artificial intelligence (XAI) elucidates the decision-making process of complex AI models and is important in building trust in model predictions. XAI explanations themselves require evaluation as to accuracy and reasonableness and in the context of use of the underlying AI model. This review details the evaluation of XAI in cardiac AI applications and has found that, of the studies examined, 37% evaluated XAI quality using literature results, 11% used clinicians as domain-experts, 11% used proxies or statistical analysis, with the remaining 43% not assessing the XAI used at all. We aim to inspire additional studies within healthcare, urging researchers not only to apply XAI methods but to systematically assess the resulting explanations, as a step towards developing trustworthy and safe models.

This paper reports a novel design of compact tuneable resonance filter with a highly extinguished and ultra-broad out-of-band rejection in CMOS compatible silicon photonics technology platform. The proposed device is designed with two identically apodized distributed grating structures for guided Fabry-Pérot resonant transmissions in a silicon on insulator rib waveguide structure. The device design parameters are optimized by theoretical simulation for a low insertion loss singly-resonant transmission peak at a desired wavelength.  However, the devices were fabricated (using in-house facilities) to demonstrate multiple resonant transmission peaks along with a singly-resonant one.  We observed that a device length of as low as ∼35 ð?œ‡m exhibits a rejection band as large as ∼60 nm with an extinction of ∼40 dB with respect to the resonant wavelength peak at ð?œ†ð?‘Ÿâˆ¼1550 nm (FWHM ∼80 pm, IL∼2 dB). The experimental results have been shown to be closely matching to our theoretical simulation and modelling results. As expected from the theoretical prediction, the trend pertaining to the trade-off between passive insertion loss and Q-value of the resonances has been observed depending on the device parameters. The thermo-optic tuning characteristics of resonant wavelengths have been obtained by integrating microheaters in the cavity. The resonance peak has been tuned at a rate of 96 pm per mW of consumed thermal power. The thermo-optic switching response has been measured to be in the order of ~5 ð?œ‡s. As a potential application, noise associated with an amplified pump wavelength (ð?œ†ð?‘ƒâˆ¼1550 nm) has been shown to be suppressed by ∼15 dB (up to the detector noise floor) which can be investigated further for large-scale integrated quantum photonic circuits. The demonstrated device can also be explored further for many other applications such as modulation, add-drop multiplexing, sensing etc.Â

Depression severity can be classified into distinct phases based on the Beck depression inventory (BDI) test scores, a subjective questionnaire. However, quantitative assessment of depression may be attained through the examination and categorization of electroencephalography (EEG) signals. Spiking neural networks (SNNs), as the third generation of neural networks, incorporate biologically realistic algorithms, making them ideal for mimicking internal brain activities while processing EEG signals. This study introduces a novel framework that for the first time, combines an SNN architecture and a long short-term memory (LSTM) structure to model the brainâ\euro™s underlying structures during different stages of depression and effectively classify individual depression levels using raw EEG signals. By employing a brain-inspired SNN model, our research provides fresh perspectives and advances knowledge of the neurological mechanisms underlying different levels of depression. The methodology employed in this study includes the utilization of the synaptic time dependent plasticity (STDP) learning rule within a 3-dimensional braintemplate structured SNN model. Furthermore, it encompasses the tasks of classifying and predicting individual outcomes, visually representing the structural alterations in the brain linked to the anticipated outcomes, and offering interpretations of the findings. Notably, our method achieves exceptional accuracy in classification, with average rates of 98% and 96% for eyes-closed and eyes-open states, respectively. These results significantly outperform state-of-the-art deep learning methods.

Paper submitted to PSCC 2024. In this paper, we present a standard power-hardware-in-the-loop testing platform that can real-time simulate detailed air-source heat pump dynamics based on well-established modeling knowledge. Distributed air-source heat pumps can be potentially used for fast frequency response. However, using air-source heat pumps for rapid modulation cannot be based on the current assumptions of linear speed-power transient characteristic, which was developed for low-speed temperature control applications. Customized setups with experimental validation options are needed to design the new fast frequency response compatible heat pump controllers. Most power system laboratories struggle in building a customized air-source heat pump, which hinders the research progress. The proposed platform is relatively universal, can fit to different heat pumps with minor modifications and be implemented on standard PHIL emulators in power system laboratories. Emulation results, real-time implementation details and model complexity metrics are presented, to assist transference of the setup to other laboratoriesÂ

Standard signal processing approaches for scintillation detectors in positron emission tomography (PET) derive accurate estimates for 511 keV photon time of interaction and energy imparted to the detection media from aggregate characteristics of electronic pulse shapes. The ultimate realization of a scintillation detector for PET is one that provides a unique timestamp and position for each detected scintillation photon. Detectors with these capabilities enable advanced concepts for three-dimensional (3D) position and time of interaction estimation with methods that exploit the spatiotemporal arrival time kinetics of individual scintillation photons. In this work, we show that taking into consideration the temporal photon emission density of a scintillator, the channel density of an analog silicon photomultiplier (SiPM) array, and employing fast electronic readout with digital signal processing, a detector that counts and timestamps scintillation photons can be realized. To demonstrate this approach, a prototype detector was constructed, comprising multichannel electronic readout for a bismuth germanate (BGO) scintillator coupled to a 4x4 SiPM array. In proof-of-concept measurements with this detector configuration, we are able to count and provide a timestamp for all optical photons produced by 511 keV photoelectric interactions. We show that this photon counting detector concept can implement 3D positioning of 511 keV photon interactions and thereby enable advanced corrections for time of interaction estimators. We outline the methodology, readout, and approach for achieving this detector capability in first-ever, proof-of-concept measurements for scintillation photon counting detector with analog silicon photomultipliers.

The calculation of Tumor Stroma Ratio (TSR) is a challenging medical issue that could improve predictions of neoadjuvant chemotherapy benefits and patient prognoses. Although several studies on breast cancer and deep learning methods have achieved promising results, the drawbacks that pixel-level semantic segmentation processes could not extract core tumor regions containing both tumor pixels and stroma pixels make it difficult to accurately calculate TSR. In this paper, we propose a Vague-Segment Technique (VST) consisting of a designed SwinV2UNet module and a modified Suzuki algorithm. Specifically, the SwinV2UNet identifies tumor pixels and generate pixel-level classification results, based on which the modified Suzuki algorithm extracts the contour of core tumor regions in terms of cosine angle. Through this way, VST obtains vaguely segmentation results of core tumor regions containing both tumor pixels and stroma pixels, where the TSR could be calculated by the formula of Intersection over Union (IOU). For the training and evaluation, we utilize the well-known The Cancer Genome Atlas (TCGA) database to create an annotated dataset, while 150 images with TSR annotations from real cases are also collected. The experimental results illustrate that the proposed VST could generate better tumor identification results compared with state-of-the-art methods, where the extracted core tumor regions lead to more consistencies of calculated TSR with senior experts compared to junior pathologists. The experimental results demonstrate the superiority of our proposed pipeline, which has promise for future clinical application.

To meet the unprecedented mobile traffic demands of future wireless networks, a paradigm shift from conventional cellular networks to distributed communication systems is imperative. Cell-free massive multiple-input multiple-output (CF-mMIMO) represents a practical and scalable embodiment of distributed/network MIMO systems. It inherits not only the key benefits of co-located massive MIMO systems but also the macro-diversity gains from distributed systems. This innovative architecture has demonstrated significant potential in enhancing network performance from various perspectives, outperforming co-located mMIMO and conventional small-cell systems. Moreover, CF-mMIMO offers flexibility in integration with emerging wireless technologies such as full-duplex (FD), non-orthogonal transmission schemes, millimeter-wave (mmWave) communications, ultra-reliable low-latency communication (URLLC),  unmanned aerial vehicle (UAV)-aided communication, and  reconfigurable intelligent surfaces (RISs). In this paper, we provide an overview of current research efforts on CF-mMIMO systems and their promising future application scenarios. We then elaborate on new requirements for CF-mMIMO networks in the context of these technological breakthroughs. We also present several current open challenges and outline future research directions aimed at fully realizing the potential of CF mMIMO systems in meeting the evolving demands of future wireless networks.

The sensor suite for assisted and automated driving functions vehicle is critical to the function of a vehicle, but also the first and most important limitation to the level of automation that the system can achieve. The advancement of 4D RADARs, providing better resolution in both azimuth and elevation compared to traditional RADAR, can assist to achieve more robust situational awareness, whilst also providing more data for perception algorithms and sensor fusion. However, like all perception sensors, 4D RADAR is also affected by numerous noise factors. To explore the sources of noise, this work identifies, classifies, and analyses automotive 4D RADAR noise factors. Overall, 22 noise factors have been considered, in combination with their effect on six 4D RADAR outputs. Finally, this work also presents and applies, for the first time, a dissimilarity metric to collected 4D RADAR data in the presence of rain with different intensities. The proposed metric is used to assess the effect of noise on the variability of the measured data, in addition it can be used to compare any 4D RADAR data. The metric, combined with other pointcloud evaluations, shows that as rainrate intensifies, the size of the pointcloud decreases, but also the variation in the measurements increases. This work presents the importance of evaluating, companding, and quantifying noise for 4D RADAR, and can pave the way for more in depth analysis of its modelling and testing of 4D RADAR for assisted and automated driving functions.

The work introduces the directional modulation transfer function (MTF) to improve the estimate of direction wave spectra from radar images of ocean waves. In contrast to the established MTF, the presented MTF better captures the directional distribution of wave energy when the spreading is high.

The decomposition of neurophysiological recordings into their constituent neural sources is of major importance to a diverse range of neuroscientific fields and neuroengineering applications. The advent of high density electrode probes and arrays has driven a major need for novel semi-automated and automated blind source separation methodologies that take advantage of the increased spatial resolution and coverage these new devices offer. Independent component analysis (ICA) offers a principled theoretical framework for such algorithms, but implementation inefficiencies often drive poor performance in practice, particularly for sparse sources. Here we observe that the use of a single non-linear optimization function to identify spiking sources with ICA often has a detrimental effect that precludes the recovery and correct separation of all spiking sources in the signal. We go on to propose a projection-pursuit ICA algorithm designed specifically for spiking sources, which uses a particle swarm methodology to adaptively traverse a polynomial family of non-linearities approximating the asymmetric cumulants of the sources. We robustly prove state-of-the-art decomposition performance on recordings from high density intramuscular probes and demonstrate how the particle swarm quickly finds optimal contrast non-linearities across a range of neurophysiological datasets.

The rapid rate of transformation in the power sector of India has placed a significant emphasis on robust grids and distributed generation units. The observable shift in the energy sector, especially in wind and solar energy, also requires smooth integration of Distributed Generation units with the existing power grid. Precise wind power generation forecast, therefore, becomes an important and complex task for the strategic planning and management of the systems. We, thus, aim towards a system that can actually provide precise wind power forecasts by applying machine learning techniques. This work proposes a comparative and comprehensive analysis of Artificial Neural Network (ANN), Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Autoregressive Integrated Moving Average (ARIMA) model. The experimentations and modelling are performed considering meteorological and historical power generation data. The study is concentrated in Kutch, Gujarat and is validated on the data collected from the Central Electricity Authority (CEA), India for power generation data and weather data collected from regional weather centres. The findings show that ARIMA outperforms the other models for non-linear data in multivariate analysis, with a MAPE score of 5.87 on the prediction dataset.

Radar is widely used in aviation, meteorology and military fields, and radar pulse signal detection has become an indispensable and essential function of cognitive radio systems as well as electronic warfare systems. In this paper, we propose a deep learning based radar signal detection method. Firstly, we propose a detection method based on raw in-phase and quadrature (IQ) input, which utilizes a convolutional neural network (CNN) to automatically learn the features of radar pulse signals and noises, so as to accomplish the detection task. To further reduce the computational complexity, we also propose a hybrid detection method that combines compressed sensing (CS) and deep learning, which reduces the length of signal by compressed downsampling, and then feeds the compressed signal to the CNN for detection. Extensive simulation results show that our proposed IQ-based method outperforms the traditional short-time Fourier transform method as well as three existing deep learning-based detection methods in terms of probability of detection. Furthermore, our proposed IQ-CS-based method is able to achieve satisfactory detection performance with significantly reduced computational complexity.

Abstractâ\euro” Wearable EEG enables us to capture large amounts of high-quality sleep data for diagnostic purposes. To make full use of this capacity we need high-performance automatic sleep scoring models. To this end, it has been noted that domain mismatch between recording equipment can be considerable, e.g. PSG to wearable EEG, but a previously observed benefit from personalizing models to individual subjects further indicates a personal domain in sleep EEG. In this work, we have investigated the extent of such a personal domain in wearable EEG, and review supervised and unsupervised approaches to personalization as found in the literature. We investigated the personalization effect of the unsupervised Adversarial Domain Adaptation and implemented an unsupervised method based on statistics alignment. No beneficial personalization effect was observed using these unsupervised methods. We find that supervised personalization leads to a substantial performance improvement on the target subject ranging from 15% Cohenâ\euro™s Kappa for subjects with poor performance (κ < 0.70) to roughly 2% on subjects with high performance (κ > 0.80). This improvement was present for models trained on both small and large datasets, indicating that even high-performance models benefit from supervised personalization. We found that this  personalization can be beneficially regularized using Kullback-Leibler regularization, leading to lower variance with negligible cost to improvement. Based on the experiments, we recommend model personalization using Kullback-Leibler regularization.

The combination of Deep Learning techniques and Raman spectroscopy shows great potential offering precise and prompt identification of pathogenic bacteria in clinical settings. However, the traditional closed-set classification approaches assume that all test samples belong to one of the known pathogens, and their applicability is limited since the clinical environment is inherently unpredictable and dynamic, unknown or emerging pathogens may not be included in the available catalogs. We demonstrate that the current state-of-the-art Neural Networks identifying pathogens through Raman spectra are vulnerable to unknown inputs, resulting in an uncontrollable false positive rate. To address this issue, first, we developed a novel ensemble of ResNet architectures combined with the attention mechanism which outperforms existing closed-world methods, achieving an accuracy of 87.8±0.1% compared to the best available modelâ\euro™s accuracy of 86.7 ± 0.4%. Second, through the integration of feature regularization by the Objectosphere loss function, our model achieves both high accuracy in identifying known pathogens from the catalog and effectively separates unknown samples drastically reducing the false positive rate. Finally, the proposed feature regularization method during training significantly enhances the performance of out-ofdistribution detectors during the inference phase improving the reliability of the detection of unknown classes. Our novel algorithm for Raman spectroscopy enables the detection of unknown, uncatalogued, and emerging pathogens providing the flexibility to adapt to future pathogens that may emerge, and has the potential to improve the reliability of Raman-based solutions in dynamic operating environments where accuracy is critical, such as public safety applications.Â

This paper explores the experimental localization of single and multiple ground RF transmitters using both traditional localization and machine learning algorithms. For the localization of a single transmitter, the setup is evaluated in two unlicensed frequency bands with and without interference. A threshold approach is proposed to improve accuracy in the presence of interference. To localize multiple transmitters, the RSSI data are divided into clusters by a k-means clustering algorithm and fed into a localization algorithm. These experimental results are preceded by an analysis phase where the UAV flight path and data collection are simulated using the QuaDRiGa channel model.

 This paper describes the development and deployment of a new type of wave sensor, termed the ΔF2luID (pronounced delta fluid), for detailed measurements of a wave’s profile during wave breaking. A large-scale laboratory measurement campaign was carried out to characterize the shoaling and breaking of waves in shallow water in a model-scale surf environment. Experiments were conducted in the Hydraulic Wave Basin Facility at the University of Iowa, with a compound planar beach installed opposite the wavemakers, which was used to produce eight wave conditions spanning deep water wavelengths from 2 to 6.5 meters (periods of 1.1 to 2 seconds) and wave amplitudes of 6.5 to 18 centimeters. The ΔF2luID sensors were co-located with ultrasonic wave gauges and arrayed across the shoaling and breaking regions of the wave field, and data from the two sensor types were hybridized to produce wave profiles that accurately captured the steepening wave profile, the overturning or spilling wave face, and the receding waterline on the wave back. A wave-by-wave statistical analysis is presented, which shows that the wave field achieves stationarity within a short duration after starting the wavemaker, and that wave heights follow an approximately normal distribution. Within the stationary process, median wave heights show a local maximum that corresponds to the observed breaker line location, after which median wave heights quickly diminish and wave height variance grows considerably. Ensemble mean wave profiles clearly show the evolution of the wave profiles across the shoaling region, with the transformation from a nonlinear, sharply-peaked wave to an overturning/plunging breaker, followed by a bore-like spilling profile. Ensemble wave profiles are used to quantify the wave set-up and the approximate normalized energy flux along the cross-shore direction, showing a gradual rise in mean water level as waves approached the shore, which grow with wavelength and wave height. Energy dissipation was evident as the waves passed through the shoaling and breaking regions, with a much more gradual rate of dissipation observed for shorter and shallower waves. A preliminary parameterization of the face profile of a breaking wave produces a good agreement with existing theoretical models, which hold promise for parameterizing wave profiles across the surf zone using non-optical measurement techniques.Â

This letter provides a simulation study on calibration of a 3D LiDAR with respect to a camera, with point-topoint and point-on-plane correspondences for calibration constraints, by extending our previously proposed 2D LiDAR calibration methods. The calibration performance is examined in terms of vertical scan view and resolution, especially in narrow/sparse scan space, in the presence of sensor noise. It contributes to a guidance on a use of various sensors and dataset depending on contexts and applications with different scan view and resolution, and an understanding of the effect of the different constraints on calibration. As a result, the combination of the point and plane constraints yields better performance than either of them.

Coded aperture snapshot spectral imager (CASSI) can recover three-dimensional hyperspectral images (HSIs) from two-dimensional compressive measurements. Recently, deep unfolding approaches were shown impressive reconstruction performance among various algorithms. Existing deep unfolding methods usually employ linear projection methods to guide the iterative learning process. However, the linear projections do not include trainable parameters and ignore the essential characteristics of HSI. This paper proposes a novel learning-based deep estimation-enhancement unfolding (DEEU) framework to improve the HSI reconstruction. The deep estimation-enhancement (DEE) module is used to guide the iterative learning process of the network based on the prior information of the CASSI system, and then exploits the intrinsic features of the reconstructed HSI in both spectral and spatial dimensions. In addition, a multi-prior ensemble learning module is proposed to further improve the reconstruction performance without increasing runtime. As with most of deep unfolding methods, we plug a convolutional neural network as a denoiser in each stage of the DEEU framework, which finally forms the proposed DEEU-Net. Comprehensive experiments on both simulation and real datasets demonstrate that the effectiveness of our DEEU framework, and our DEEU-Net can achieve both high reconstruction quality and speed, outperforming the state-of-the-art methods.

In wireless communication systems, the asynchronization of the oscillators in the transmitter and the receiver along with the Doppler shift due to relative movement may lead to the presence of carrier frequency offset (CFO) in the received signals. Estimation of CFO is crucial for subsequent processing such as coherent demodulation. In this brief, we demonstrate the utilization of deep learning for CFO estimation by employing a residual network (ResNet) to learn and extract signal features from the raw in-phase (I) and quadrature (Q) components of the signals. We use multiple modulation schemes in the training set to make the trained model adaptable to multiple modulations or even new signals. In comparison to the commonly used traditional CFO estimation methods, our proposed IQ-ResNet method exhibits superior performance across various scenarios including different oversampling ratios, various signal lengths, and different channels.

Co-Channel Interference is a classical problem in cellular systems that has been studied extensively and several methods have been proposed to overcome it. These include interference rejection techniques as well as joint detection techniques. We have previously proposed a joint detection technique for MSK-type signals that works quite well in certain conditions. In this paper, we formally present what we call Orthogonal MSK and postulate that if two MSK signals have a 90-degree phase offset between them then both can be detected successfully increasing the spectral efficiency two-fold. This technique works well even if the two signals are near equal power and have the same carrier frequency.

Continuous Hearth Rate (HR) monitoring based on photoplethysmography (PPG) sensors is a crucial feature of almost all wrist-worn devices. However, arm movements lead to the creation of Motion Artifacts (MA), affecting the performance of PPG-based HR tracking. This problem is commonly tackled by exploiting the recorded accelerometer data to correlate them with the PPG signal and eventually clean it. Thus, automatic fusion techniques based on Deep Learning (DL) algorithms have been proposed, but they are considered too large and complex to be deployed on wearable devices. The current work presents a novel and lightweight DL architecture, PULSE, comprised of temporal convolutions and feature-level multi-head cross-attention to improve sensor fusion’s effectiveness. Moreover, we propose a relation-based knowledge distillation mechanism to pass PULSE’s knowledge to a student network that utilizes modality-wise convolutions to replace the attention module and mimic the teacher’s performance with 5x fewer parameters. The teacher and student are evaluated on the most extensive available dataset, PPG-DaLiA, with PULSE reducing the mean absolute error by 8.2% compared to the best state-of-the-art model while simultaneously reducing the inference latency by 1.6x. The student model is further compressed using post-training quantization and deployed on two microcontrollers, demonstrating its suitability for real-time execution, having a close-to-state-of-the-art MAE of 4.81 BPM (+0.40 BPM), but a 10.9x lower memory footprint of 37.9 kB, and consuming 45.9x lower energy (0.577 mJ).

Most single image super resolution (SISR) methods are developed on synthetic low resolution (LR) and  high resolution (HR) image pairs, which are simulated by a  predetermined degradation operation, such as bicubic  downsampling. However, these methods only learn the  inverse process of the predetermined operation, which fails to super resolve the real-world LR images, whose true  formulation deviates from the predetermined operation. To  address this, we propose a novel SR framework named  hardware-aware super-resolution (HASR) network that first extracts hardware information, particularly the camera degradation information. The LR images are then super resolved by integrating the extracted information. To  evaluate the performance of HASR network, we build a dataset named Real-Micron from real-world micron-scale  patterns. The paired LR and HR images are captured by  changing the objectives and registered using a developed  registration algorithm. Transfer learning is implemented  during the training of Real-Micron dataset due to the lack  of amount of data. Experiments demonstrate that by  integrating the degradation information, our proposed  network achieves state-of-the-art performance for the blind  SR task on both synthetic and real-world datasets. Impact Statementâ\euro” The proposed HASR method has  significant impact on various areas, such as enhancing the  accurate inspection of manufactured products for quality  control and enhancing the resolution of medical images to  enable more accurate diagnosis and healthcare. Current SR solutions neglect the uniqueness of each imaging system,  hence cannot produce accurate HR images across the  different systems. Taking advantage of the known hardware information, HASR can differentiate low?resolution images across different imaging systems and  produce HR images that are closer to the real-world  scenario. Given sufficient training images, the proposed  HASR method can overcome the physical optical limitation  and generate higher quality images. The proposed method  improves the overall performance by about 0.2 dB and 0.5  dB on the synthetic and the real-world datasets,  respectively. Â

A document by Ci Lin . Click on the document to view its contents.