Data-driven Filters Research Articles

Insights Into Deep Non-Linear Filters for Improved Multi-Channel Speech Enhancement

The key advantage of using multiple microphones for speech enhancement is that spatial filtering can be used to complement the tempo-spectral processing. In a traditional setting, linear spatial filtering (beamforming) and single-channel post-filtering are commonly performed separately. In contrast, there is a trend towards employing deep neural networks (DNNs) to learn a joint spatial and tempo-spectral non-linear filter, which means that the restriction of a linear processing model and that of a separate processing of spatial and tempo-spectral information can potentially be overcome. However, the internal mechanisms that lead to good performance of such data-driven filters for multi-channel speech enhancement are not well understood. Therefore, in this work, we analyse the properties of a non-linear spatial filter realized by a DNN as well as its interdependency with temporal and spectral processing by carefully controlling the information sources (spatial, spectral, and temporal) available to the network. We confirm the superiority of a non-linear spatial processing model, which outperforms an oracle linear spatial filter in a challenging speaker extraction scenario for a low number of microphones by 0.24 POLQA score. Our analyses reveal that in particular spectral information should be processed jointly with spatial information as this increases the spatial selectivity of the filter. Our systematic evaluation then leads to a simple network architecture, that outperforms state-of-the-art network architectures on a speaker extraction task by 0.22 POLQA score and by 0.32 POLQA score on the CHiME3 data.

MDFNet: an unsupervised lightweight network for ear print recognition.

In this paper, we propose an unsupervised lightweight network with a single layer for ear print recognition. We refer to this method by MDFNet because it relies on gradient Magnitude and Direction alongside with responses of data-driven Filters. At first, we align ear using Convolution Neural Network (CNN) and Principal Component Analysis (PCA). MDFNet starts by generating a filter bank from training images using PCA. This is followed by a twofold layer, which comprises two operations namely convolution using learned filters and computation of gradient image. To prevent over-fitting, a binary hashing process is done by combining different filter responses into a single feature map. Then, we separately construct histograms for each of gradient magnitude and direction according to the feature map. These histograms are then normalized, using power-L2 rule, to cope with illumination disparity. Several fusion rules are evaluated to combine the two histograms. The main novelty of MDFNet lies in its simple architecture, effectiveness, the good compromise between processing time and performance it provides along with its high robustness to occlusion. We conduct extensive experiments on three public datasets namely AWE, AMI and IIT Delhi II. Experimental results demonstrate the effectiveness of MDFNet, which achieves high recognition rates (82.5%, 97.67% and 98.96%, respectively), and outperformed several state of the art methods with a high robustness to occlusion. Experiments revealed also the actual need for considering ear alignment.

Epileptic Spike Detection Using Neural Networks With Linear-Phase Convolutions

To cope with the lack of highly skilled professionals, machine learning with proper signal processing is key for establishing automated diagnostic-aid technologies with which to conduct epileptic electroencephalogram (EEG) testing. In particular, frequency filtering with the appropriate passbands is essential for enhancing the biomarkers-such as epileptic spike waves-that are noted in the EEG. This paper introduces a novel class of neural networks (NNs) that have a bank of linear-phase finite impulse response filters at the first layer as a preprocessor that can behave as bandpass filters that extract biomarkers without destroying waveforms because of a linear-phase condition. Besides, the parameters of the filters are also data-driven. The proposed NNs were trained with a large amount of clinical EEG data, including 15833 epileptic spike waveforms recorded from 50 patients, and their labels were annotated by specialists. In the experiments, we compared three scenarios for the first layer: no preprocessing, discrete wavelet transform, and the proposed data-driven filters. The experimental results show that the trained data-driven filter bank with supervised learning behaves like multiple bandpass filters. In particular, the trained filter passed a frequency band of approximately 10-30 Hz. Moreover, the proposed method detected epileptic spikes, with the area under the receiver operating characteristic curve of 0.967 in the mean of 50 intersubject validations.

Grouped variable selection for generalized eigenvalue problems

Many problems require the selection of a subset of variables from a full set of optimization variables. The computational complexity of an exhaustive search over all possible subsets of variables is, however, prohibitively expensive, necessitating more efficient but potentially suboptimal search strategies. We focus on sparse variable selection for generalized Rayleigh quotient optimization and generalized eigenvalue problems. Such problems often arise in the signal processing field, e.g., in the design of optimal data-driven filters. We extend and generalize existing work on convex optimization-based variable selection using semidefinite relaxations toward group-sparse variable selection using the ℓ1,∞-norm. This group-sparsity allows, for instance, to perform sensor selection for spatio-temporal (instead of purely spatial) filters, and to select variables based on multiple generalized eigenvectors instead of only the dominant one. Furthermore, we extensively compare our method to state-of-the-art methods for sensor selection for spatio-temporal filter design in a simulated sensor network setting. The results show both the proposed algorithm and backward greedy selection method best approximate the exhaustive solution. However, the backward greedy selection has more specific failure cases, in particular for ill-conditioned covariance matrices. As such, the proposed algorithm is the most robust currently available method for group-sparse variable selection in generalized eigenvalue problems.

Deep neural architectures for dialect classification with single frequency filtering and zero-time windowing feature representations.

The goal of this study is to investigate advanced signal processing approaches [single frequency filtering (SFF) and zero-time windowing (ZTW)] with modern deep neural networks (DNNs) [convolution neural networks (CNNs), temporal convolution neural networks (TCN), time-delay neural network (TDNN), and emphasized channel attention, propagation and aggregation in TDNN (ECAPA-TDNN)] for dialect classification of major dialects of English. Previous studies indicated that SFF and ZTW methods provide higher spectro-temporal resolution. To capture the intrinsic variations in articulations among dialects, four feature representations [spectrogram (SPEC), cepstral coefficients, mel filter-bank energies, and mel-frequency cepstral coefficients (MFCCs)] are derived from SFF and ZTW methods. Experiments with and without data augmentation using CNN classifiers revealed that the proposed features performed better than baseline short-time Fourier transform (STFT)-based features on the UT-Podcast database [Hansen, J. H., and Liu, G. (2016). "Unsupervised accent classification for deep data fusion of accent and language information," Speech Commun. 78, 19-33]. Even without data augmentation, all the proposed features showed an approximate improvement of 15%-20% (relative) over best baseline (SPEC-STFT) feature. TCN, TDNN, and ECAPA-TDNN classifiers that capture wider temporal context further improved the performance for many of the proposed and baseline features. Among all the baseline and proposed features, the best performance is achieved with single frequency filtered cepstral coefficients for TCN (81.30%), TDNN (81.53%), and ECAPA-TDNN (85.48%). An investigation of data-driven filters, instead of fixed mel-scale, improved the performance by 2.8% and 1.4% (relatively) for SPEC-STFT and SPEC-SFF, and nearly equal for SPEC-ZTW. To assist related work, we have made the code available ([Kethireddy, R., and Kadiri, S. R. (2022). "Deep neural architectures for dialect classification with single frequency filtering and zero-time windowing feature representations," https://github.com/r39ashmi/e2e_dialect (Last viewed 21 December 2021)].).

TOWARD DATA-DRIVEN FILTERS IN PARAVIEW

Recent progress in scientific visualization has expanded the scope of visualization from being merely a way of presentation to an analysis and discovery tool. A given visualization result is usually generated by applying a series of transformations or filters to the underlying data. Nowadays, such filters use deterministic algorithms to process the data. In this work, we aim at extending this methodology towards data-driven filters, thus filters that expose the abilities of pre-trained machine learning models to the visualization system. The use of such data-driven filters is of particular interest in fields like segmentation, classification, etc., where machine learning models regularly outperform existing algorithmic approaches. To showcase this idea, we couple Paraview, the well-known flow visualization tool, with PyTorch, a deep learning framework. Paraview is extended by plugins that allow users to load pre-trained models of their choice in the form of newly developed filters. The filters transform the input data by feeding it into the model and then provide the model's output as input to the remaining visualization pipeline. A series of simplistic use cases for segmentation and classification on image and fluid data is presented to showcase the technical applicability of such data-driven transformations in Paraview for future complex analysis tasks.

Recursive Singular Spectrum Analysis for Induction Machines Unbalanced Rotor Fault Diagnosis

One of the major challenges of diagnosing rotor symmetry faults in induction machines (IMs) is severe modulation of fault and supply frequency components. In particular, existing techniques are not able to identify fault components in the case of low slips. In this article, this problem is tackled by proposing a novel approach. First, a new use of singular spectrum analysis (SSA), as a powerful spectrum analyzer, is introduced for fault detection. Our idea is to treat the stator current signature of the wound rotor IM as a time series. In this approach, the current signature is decomposed into several eigenvalue spectra (rather than frequency spectra) to find a subspace where the fault component is recognizable. Subsequently, the fault component is detected using some data-driven filters constructed with the knowledge about characteristics of supply and fault components. Then, an inexpensive peak localization procedure is applied to the power spectrum of the fault component to identify the exact frequency of the fault. The fault detection and localization methods are then combined in a recursive regime to further improve the diagnosis’ performance particularly at high rotor speeds and small rotor faults. The proposed approach is data-driven and is directly applied to the raw signal with no suppression or filtration of the frequency harmonics with a low computational complexity. The numerical results obtained with real data at several rotation speeds and fault severities unveil the effectiveness and real-time feature of the proposed approach.

TomahaqCompanion: A Tool for the Creation and Analysis of Isobaric Label Based Multiplexed Targeted Assays.

Triggered by Offset, Multiplexed, Accurate mass, High resolution, and Absolute Quantitation (TOMAHAQ) is a recently introduced targeted proteomics method that combines peptide and sample multiplexing. TOMAHAQ assays enable sensitive and accurate multiplexed quantification by implementing an intricate data collection scheme that comprises multiple MSn scans, mass inclusion lists, and data-driven filters. Consequently, manual creation of TOMAHAQ methods can be time-consuming and error prone, while the resulting TOMAHAQ data may not be compatible with common mass spectrometry analysis pipelines. To address these concerns we introduce TomahaqCompanion, an open-source desktop application that enables rapid creation of TOMAHAQ methods and analysis of TOMAHAQ data. Starting from a list of peptide sequences, a user can perform each step of TOMAHAQ assay development including (1) generation of priming run target list, (2) analysis of priming run data, (3) generation of TOMAHAQ method file, and (4) analysis and export of quantitative TOMAHAQ data. We demonstrate the flexibility of TomahaqCompanion by creating a variety of methods testing TOMAHAQ parameters (e.g., number of SPS notches, run length, etc.). Lastly, we analyze an interference sample comprising heavy yeast peptides, a standard human peptide mixture, TMT11-plex, and super heavy TMT (shTMT) isobaric labels to demonstrate ∼10-200 attomol limit of quantification within a complex background using TOMAHAQ.

An Integrated Data Quality System for Species Observations

Species-level observational data comprise the largest and fastest-growing part of the Global Biodiversity Information Facility (GBIF). The largest single contributor of species observations is eBird, which so far has contributed more than 361 million records to GBIF. eBird engages a vast network of human observers (citizen-scientists) to report bird observations, with the goal of estimating the range, abundance, habitat preferences, and trends of bird species at high spatial and temporal resolutions across each species’ entire life-cycle. Since its inception, eBird has focused on improving the data quality of its observations, primarily focused in two areas: ensuring that participants describe how they gathered their observations and, all observations are reviewed for accuracy. In this presentation I will review how this is done in eBird. Standardized Data Collection. eBird gathers bird observations based on how bird watchers typically observe birds with units of data collection being “checklists” of zero or more species including a count of individuals for each species observed. Participants choose the location where they made their observations and submit their checklists via Mobile Apps (50% of all submissions) or the website (50% of all submissions). All checklists are submitted in a standard format identifying where, how, and with whom they made their observations. Mobile apps precisely record locations, the track taken, and the distance they traveled while making the observations. The start time and duration of surveys are also recorded. All observers must report whether they reported all the birds they detected and identified, which allows analysts to infer absence of birds if they were not reported. All data are stored within an Oracle data management framework. Data Accuracy. The most significant data quality challenge for species observations is detecting and correctly identifying organisms to species. The issue involves how to handle both false positives — the misidentification of an observed organism, and false negatives—failing to report a species that was present. The most egregious false positives can be identified as anomalies that fall outside the norm of occurrence for a species at a particular time or space. However, false positives can also be misidentifications of common species. These challenges are addressed by: Data-driven filters. eBird’s existing data can identify and flag potentially erroneous records at increasingly fine spatial, temporal, and user-specific scales. These filters can identify outliers and likely errors, which are the foundation of the eBird review process. By using the vetted data to identify outliers, data quality checks run against expected occurrence probabilities at very fine scales and identify anomalies during data submission (including on mobile devices). Incorporate observer expertise scores. Observer differences are the largest source of variability in eBird data. Assessment of observer metrics, and the inclusion of these data in species distribution models, improves analysis output and model performance. Expert reviewer network. More than 2000 volunteers review records identified by the data-driven filters and contact data submitters to confirm their observations. The existing data quality process functions globally. Currently the approach is focused on misidentified birds, but in the future will also involve collection event issues (e.g., issues with protocol, location, or methodology), sensitive species, exotic species, and better handle widely-observed individual rarities. Additional tools are also to be developed to help editors improve efficiency and better prioritize review. In 2017, 4,107,757 observations representing 4.6% of all eBird records submitted were flagged for review by the data driven filters. Of these records 57.4% were validated and 42.6% were invalidated.

P186. Correlates to influence user performance in a hand motor rehabilitation task

Introduction The power of oscillatory EEG signal components has been related to attention or mental workload. Thus it may allow to decode and track mental states related to the performance of a user on a specific task. In motor rehabilitation after stroke, knowledge about suitable or impeding brain states may be informative enough to causally influence the training performance. A prerequisite for such real-time manipulation is the ability to decode ongoing EEG signals. Solutions are provided by the field of Brain–Computer Interface (BCI) systems. We propose the application of BCI methods in the context of decoding mental states relevant for successful stroke motor rehabilitation training. Extracting oscillatory sources from a pre-trial EEG interval of the alpha band with optimized data-driven filters, we show that the sources are informative about trial-wise reaction time (RT). Methods A paradigm to assess and train hand motor function in acute and chronic stroke patients is the Sequential Visual Isometric Pinch Task (SVIPT, Reis et al., 2009). Current studies have shown that motor improvements achieved using SVIPT training will generalize to a range of hand-motor tasks. We report on simulated online results of a novel EEG-tracked SVIPT system. Specifically, we explored the usefulness of alpha band (8–14 Hz) pre-trial information (−800 to 0 ms) to predict fluctuations of subject’s reaction times (RT) in SVIPT. Data was collected from three subjects (age 23, 25, 33) using 63 EEG channels plus 1 EOG channel. Sampled at 1 kHz, band-filtered to 0.7–90 Hz and sub-sampled at 500 Hz, the data was cleaned from eye- and muscle artifacts. Oscillatory EEG components, which co-modulate in power with RT, were optimized using the novel SPoC method (Dahne et al., 2014). All results reported were computed following a 5-fold cross-validation procedure. Results For all subjects, SPoC results in alpha sources with focus on parietal-occipital channels. Fig. 1 shows the estimated probability density distributions (pdf) of the true trial-wise RT using only the 33% of fastest and slowest trials, as predicted by the power of pre-trial SPoC sources. To derive the pdfs, RT values estimated using the power of the oscillatory component were sorted. For VP2 and VP3, a slight separation between groups of fast and slow trials is observed, with left-shifted pdfs for predicted faster trials. For data of VP1 a separation is not evident. The mean RT values dependent on the percentage of included quickest trials are plotted in Fig. 2 . An asymptotic trend is present for VP2 and VP3, which indicates, that taking only trials with the predicted shortest RT values in fact lead to shorter average RTs. Conclusion The proposed EEG-tracked SVIPT is a promising tool for the prediction of trial-wise performance. Eventually applied in an online motor rehabilitation scenario, it may be versatile to gate the start of each single trial. In the hands of therapists it may allow to make training trials either harder or easier for the subject. Acknowledgement Funded by DFG, BrainLinks-BrainTools (EXC1086).

On the design of filters for fourier and oSVD-based deconvolution in bolus tracking perfusion MRI

Bolus tracking perfusion evaluation relies on the deconvolution of a tracers concentration time-courses in an arterial and a tissue voxel following the tracer kinetic model. The object of this work is to propose a method to design a data-driven Tikhonov regularization filter in the Fourier domain and to compare it to the singular value decomposition (SVD)-based approaches using the mathematical equivalence of Fourier and circular SVD (oSVD). The adaptive filter is designed using Tikhonov regularization that depends on only one parameter. Using a simulation, such an optimal parameter that minimizes the sum of statistical and systematic error is determined as a function of the first moment difference between the tissue and the arterial curve and the contrast to noise ratios of the input data (CNR( a ) in arteries and CNR( t ) in tissue). The performance of the method is evaluated and compared to oSVD in simulations and measured data. The proposed method yields a smaller flow underestimation especially for high flows when compared to the oSVD approach with constant threshold. However, this improvement comes to the price of an increased uncertainty of the flow values. The translation of the Tikhonov regularization parameter to an adaptive oSVD-threshold is in good agreement with the literature. The proposed method is a comprehensive approach for the design of data-driven filters that can be easily adapted to specific needs.

Market Beta Dynamics and Portfolio Efficiency

This paper introduces a new estimation for the dynamics of betas. It combines two previously separate approaches in the literature, data-driven filters and parametric methods. Namely, we show how to estimate the parametric beta dynamics by instrumental variables combined with block-sampling - but not overlapping window filters - of data-driven betas. Instrumental variables are needed because of the measurement errorsin empirical betas. We find that, while betas are very strongly autocorrelated, neither aggregate nor firm-specific variables explain much of their quarterly variation. We then compare block-samplers and overlapping window filters using a criterion of economic significance. Namely, we track the out-of-sample performance of portfolios optimized subject to target beta constraints. For target betas of zero, the case of many hedge funds, we show that estimation error results in systematic overshooting of the target beta. These portfolios benefit from the use of medium to long term estimation windows of daily returns.