While the transformer has emerged as the eminent neural architecture, several independent lines of research have emerged to address its limitations. Recurrent neural approaches have observed a lot of renewed interest, including the extended long short-term memory (xLSTM) architecture, which reinvigorates the original LSTM. However, while xLSTMs have shown competitive performance compared to the transformer, their viability for learning self-supervised general-purpose audio representations has not been evaluated. This work proposes Audio xLSTM (AxLSTM), an approach for learning audio representations from masked spectrogram patches in a self-supervised setting. Pretrained on the AudioSet dataset, the proposed AxLSTM models outperform comparable self-supervised audio spectrogram transformer (SSAST) baselines by up to 25% in relative performance across a set of ten diverse downstream tasks while having up to 45% fewer parameters.

The efficiency of Bayesian optimization (BO) relies heavily on the choice of the Gaussian process (GP) kernel, which plays a central role in balancing exploration and exploitation under limited evaluation budgets. Traditional BO methods often rely on fixed or heuristic kernel selection strategies, which can result in slow convergence or suboptimal solutions when the chosen kernel is poorly suited to the underlying objective function. To address this limitation, we propose a freshly-baked Context-Aware Kernel Evolution (CAKE) to enhance BO with large language models (LLMs). Concretely, CAKE leverages LLMs as the crossover and mutation operators to adaptively generate and refine GP kernels based on the observed data throughout the optimization process. To maximize the power of CAKE, we further propose BIC-Acquisition Kernel Ranking (BAKER) to select the most effective kernel through balancing the model fit measured by the Bayesian information criterion (BIC) with the expected improvement at each iteration of BO. Extensive experiments demonstrate that our fresh CAKE-based BO method consistently outperforms established baselines across a range of real-world tasks, including hyperparameter optimization, controller tuning, and photonic chip design. Our code is publicly available at https://github.com/richardcsuwandi/cake.

Gaussian processes (GPs) stand as crucial tools in machine learning and signal processing, with their effectiveness hinging on kernel design and hyperparameter optimization. This article presents a novel GP linear multiple kernel (LMK) and a generic sparsity-aware distributed learning framework to optimize the hyperparameters. The newly proposed grid spectral mixture product (GSMP) kernel is tailored for multidimensional data, effectively reducing the number of hyperparameters while maintaining good approximation capability. We further demonstrate that the associated hyperparameter optimization of this kernel yields sparse solutions. To exploit the inherent sparsity of the solutions, we introduce the sparse linear multiple kernel learning (SLIM-KL) framework. The framework incorporates a quantized alternating direction method of multipliers (ADMMs) scheme for collaborative learning among multiple agents, where the local optimization problem is solved using a distributed successive convex approximation (DSCA) algorithm. SLIM-KL effectively manages large-scale hyperparameter optimization for the proposed kernel, simultaneously ensuring data privacy and minimizing communication costs. The theoretical analysis establishes convergence guarantees for the learning framework, while experiments on diverse datasets demonstrate the superior prediction performance and efficiency of our proposed methods.

The Gaussian process state-space model (GPSSM) has attracted extensive attention for modeling complex nonlinear dynamical systems. However, the existing GPSSM employs separate Gaussian processes (GPs) for each latent state dimension, leading to escalating computational complexity and parameter proliferation, thus posing challenges for modeling dynamical systems with high-dimensional latent states. To surmount this obstacle, we propose to integrate the efficient transformed Gaussian process (ETGP) into the GPSSM, which involves pushing a shared GP through multiple normalizing flows to efficiently model the transition function in high-dimensional latent state space. Additionally, we develop a corresponding variational inference algorithm that surpasses existing methods in terms of parameter count and computational complexity. Experimental results on diverse synthetic and real-world datasets corroborate the efficiency of the proposed method, while also demonstrating its ability to achieve similar inference performance compared to existing methods. Code is available at https://github.com/zhidilin/gpssmProj.

A fundamental problem in cognitive radio is spectrum sensing, which detects the presence of the primary users in a licensed spectrum. To boost the detection performance and robustness, the multiantenna detector has been investigated and various related methods have been developed, e.g., the energy detector, the eigenvalue arithmetic-to-geometric mean detector, and the generalized likelihood ratio test detector. Cooperative sensing, which makes use of multiple receivers distributed in different locations, has the advantage of being able to make full use of the distributed antennas and enjoy a high spatial diversity gain. However, the successful employment of cooperative sensing depends on the reliable information exchange among the cooperating receivers over a long range, which may be impractical for real-world scenarios. In this paper, we consider the scenario where each receiving node can only broadcast its received raw data in a short-range communication fashion. We propose a novel cooperative sensing scheme by allowing each node to send to the fusion center only local correlation coefficients, computed within a neighborhood. A detection algorithm, based on matrix factorization of the partially received sample covariance matrix, i.e., with missing entries, is proposed. The performance of our proposed cooperative scheme is verified via numerical experiments.

In this work, we propose a Multi-Window Masked Autoencoder (MW-MAE) fitted with a novel Multi-Window Multi-Head Attention (MW-MHA) module that facilitates the modelling of local-global interactions in every decoder transformer block through attention heads of several distinct local and global windows. Empirical results on ten downstream audio tasks show that MW-MAEs consistently outperform standard MAEs in overall performance and learn better general-purpose audio representations, along with demonstrating considerably better scaling characteristics. Investigating attention distances and entropies reveals that MW-MAE encoders learn heads with broader local and global attention. Analyzing attention head feature representations through Projection Weighted Canonical Correlation Analysis (PWCCA) shows that attention heads with the same window sizes across the decoder layers of the MW-MAE learn correlated feature representations which enables each block to independently capture local and global information, leading to a decoupled decoder feature hierarchy.

Variational autoencoders (VAE) are one of the most prominent deep generative models for learning the underlying statistical distribution of high-dimensional data. However, training VAEs suffers from a severe issue called posterior collapse; that is, the learned posterior distribution collapses to the assumed/pre-selected prior distribution. This issue limits the capacity of the learned posterior distribution to convey data information. Previous work has proposed a heuristic training scheme to mitigate this issue, in which the core idea is to train the encoder and the decoder in an alternating fashion. However, there is still no theoretical interpretation of this scheme, and this paper, for the first time, fills in this gap by inspecting the previous scheme under the lens of the expectation maximization (EM) framework. Under this framework, we propose a novel EM-type training algorithm that gives a controllable optimization process and it allows for further extensions, e.g., employing implicit distribution models. Experimental results have corroborated the superior performance of the proposed EM-type VAE training algorithm in terms of various metrics.

Extracting information from fMRI data constitutes a broad active area of research. Current techniques still present several limitations; some ignore relevant aspects regarding the brain functioning or lack of interpretability. In an effort to overcome such limitations, we introduce an extension of the sparse matrix factorization approach to a multilinear decomposition. The proposed model is built upon natural justifiable assumptions and better accommodates the brain behavior. Tests on realistic synthetic as well as real fMRI datasets demonstrate significant performance gains over existing methods of this kind.

Transfer entropy can to a certain degree assess the direction in addition to the strength of the couplings within dynamic time series. The greater the transfer entropy, the greater the strength of the dependency between time series. In this work, we are interested in quantifying the effect that a given time series (e.g., an external stimuli) has upon the coupling strength between other time series. Towards that end, we define a directed dependency index based on the difference of two causally conditioned transfer entropies. We then provide a lower bound for the dependency index, and demonstrate on synthetic data that this lower bound can be efficiently computed.

This work addresses adversarial robustness in deep learning by considering deep networks with stochastic local winner-takes-all (LWTA) activations. This type of network units result in sparse representations from each model layer, as the units are organized in blocks where only one unit generates a non-zero output. The main operating principle of the introduced units lies on stochastic arguments, as the network performs posterior sampling over competing units to select the winner. We combine these LWTA arguments with tools from the field of Bayesian non-parametrics, specifically the stick-breaking construction of the Indian Buffet Process, to allow for inferring the sub-part of each layer that is essential for modeling the data at hand. Then, inference is performed by means of stochastic variational Bayes. We perform a thorough experimental evaluation of our model using benchmark datasets. As we show, our method achieves high robustness to adversarial perturbations, with state-of-the-art performance in powerful adversarial attack schemes.

We are currently experiencing the dawn of what is known as the fourth industrial revolution. At the center of this historical happening, as one of the key enabling technologies, lies a discipline that deals with data and whose goal is to extract information and related knowledge that is hidden in it, in order to make predictions and, subsequently, take decisions. Machine learning (ML) is the name that is used as an umbrella to cover a wide range of theories, methods, algorithms, and architectures that are used to this end. The articles in this special issue cover promising developments in the related areas of machine learning and deep neural networks and offers possible paths for the future.

We propose a novel Gaussian process kernel that takes advantage of a deep neural network (DNN) structure but retains good interpretability. The resulting kernel is capable of addressing four major issues of the previous works of similar art, i.e., the optimality, explainability, model complexity, and sample efficiency. Our kernel design procedure comprises three steps: (1) Derivation of an optimal kernel with a non-stationary dot product structure that minimizes the prediction/test mean-squared-error (MSE); (2) Decomposition of this optimal kernel as a linear combination of shallow DNN subnetworks with the aid of multi-way feature interaction detection; (3) Updating the hyper-parameters of the subnetworks via an alternating rationale until convergence. The designed kernel does not sacrifice interpretability for optimality. On the contrary, each subnetwork explicitly demonstrates the interaction of a set of features in a transformation function, leading to a solid path toward explainable kernel learning. We test the proposed kernel with both synthesized and real-world data sets, and the proposed kernel is superior to its competitors in terms of prediction performance in most cases. Moreover, it tends to maintain the prediction performance and be robust to data over-fitting issue, when reducing the number of samples.

The aim of this work is to enable inference of deep networks that retain high accuracy for the least possible model complexity, with the latter deduced from the data during inference. To this end, we revisit deep networks that comprise competing linear units, as opposed to nonlinear units that do not entail any form of (local) competition. In this context, our main technical innovation consists in an inferential setup that leverages solid arguments from Bayesian nonparametrics. We infer both the needed set of connections or locally competing sets of units, as well as the required floating-point precision for storing the network parameters. Specifically, we introduce auxiliary discrete latent variables representing which initial network components are actually needed for modeling the data at hand, and perform Bayesian inference over them by imposing appropriate stick-breaking priors. As we experimentally show using benchmark datasets, our approach yields networks with less computational footprint than the state-of-the-art, and with no compromises in predictive accuracy.

We propose a modified spectral mixture (SM) kernel that serves as a universal stationary kernel for temporal Gaussian process regression (GPR). The kernel is named grid spectral mixture (GSM) kernel as we fix the frequency and variance parameters in the original SM kernel to a set of pre-selected grid points. The hyper-parameters are the non-negative weights of all sub-kernel functions and the resulting optimization task falls under the difference-of-convex programming. Due to the nice structure of the optimization problem, the hyper-parameters are solved by an efficient majorization-minimization method instead of the gradient descent methods. It turns out that the solution is sparse, which provides us with a principled guideline to identify the important frequency components of the data. Experimental results based on various classic time series data sets corroborate that the proposed GPR with GSM kernel significantly outperforms the GPR with SM kernel in terms of both the mean-squared-error (MSE) and the stability of the optimization algorithm.

Extracting information from functional magnetic resonance images (fMRI) has been a major area of research for more than two decades. The goal of this work is to present a new method for the analysis of fMRI data sets, that is capable to incorporate a priori available information, via an efficient optimization framework. Tests on synthetic data sets demonstrate significant performance gains over existing methods of this kind.

This paper proposes a novel algorithm for online distributed dictionary learning, where a set of nodes is requested to collectively estimate a common dictionary via sequentially received data vectors. At each time instance, in which a new datum becomes available, the sparse representation of the data with respect to the estimated dictionary is computed locally at each node by employing a sparsity promoting algorithm. In the sequel, the nodes cooperate in order to collaboratively update the dictionary via the distributed Recursive Least Squares (RLS) algorithm. Numerical examples, both with synthetic and real data, validate that the performance of the proposed algorithm is comparable to that of centralized state of the art algorithms.

A novel algorithm for distributed labeling of speech sources is proposed. We consider a wireless sensor network comprising devices that are equipped with multiple microphones, which can “hear” a number of speech signals. The labeling task is performed in a decentralized fashion with a new two-step approach. The first step corresponds to the distributed extraction of proper source-specific features from the mixed signals. In the second step, these features are exploited via a distributed unsupervised learning technique. We present approaches that can be used in hierarchically organized or in non-hierarchically organized network configurations. Numerical examples using real data display the performance of the proposed technique.

A promising approach when dealing with massive data sets is to apply randomized dimensionality reduction and then operate in lower dimensions. This paper deals with the randomized linear regression task in the case where the available data are sporadically corrupted. Instead of relying to minimization of norms, which are robust to outliers, an alternative route is taken. Building upon the observation that outliers can be detected, while operating in a low dimensional randomized projections produced embedding, a mechanism for iteratively detecting and excluding corrupted data is proposed. As a result, the linear regression is performed using conventional LS approximation, without the need to resort to linear programming-based ℓ1norm minimization tasks.

This paper presents pattern classification to a predefined set of classes as a missing data task. This is achieved by first augmenting the feature vector of each training pattern with the corresponding binary codeword representing its class. A Restricted Boltzmann Machine (RBM) or a Dictionary Learning (DL) algorithm is then trained on the augmented feature space. During the classification stage, the binary codeword of the unknown pattern is treated as missing data. In the case of the RBM, it is filled in by means of an alternating Gibbs sampling procedure. In the case of the DL method, the set of atoms in the dictionary is first learned from the training data, and the label of the unknown pattern is predicted based on those atoms that represent this pattern. Application of the method in an audio genre classification task verifies that the obtained results are highly competitive compared with state-of-the-art methods. Moreover, the DL approach lends itself readily for online implementations, in line with the current trend in big data applications.

The paper presents a new framework for complex support vector regression (SVR) as well as Support Vector Machines (SVM) for quaternary classification. The method exploits the notion of widely linear estimation to model the input-out relation for complex-valued data and considers two cases: 1) the complex data are split into their real and imaginary parts and a typical real kernel is employed to map the complex data to a complexified feature space and 2) a pure complex kernel is used to directly map the data to the induced complex feature space. The recently developed Wirtinger's calculus on complex reproducing kernel Hilbert spaces is employed to compute the Lagrangian and derive the dual optimization problem. As one of our major results, we prove that any complex SVM/SVR task is equivalent with solving two real SVM/SVR tasks exploiting a specific real kernel, which is generated by the chosen complex kernel. In particular, the case of pure complex kernels leads to the generation of new kernels, which have not been considered before. In the classification case, the proposed framework inherently splits the complex space into four parts. This leads naturally to solving the four class-task (quaternary classification), instead of the typical two classes of the real SVM. In turn, this rationale can be used in a multiclass problem as a split-class scenario based on four classes, as opposed to the one-versus-all method; this can lead to significant computational savings. Experiments demonstrate the effectiveness of the proposed framework for regression and classification tasks that involve complex data.

In this paper, an Adaptive Projected Subgradient Method (APSM) based algorithm for robust subspace tracking is introduced. A properly chosen cost function is constructed at each time instance and the goal is to seek for points, which belong to the zero level set of this function; i.e., the set of points which score a zero loss. In each iteration, an outlier detection mechanism is employed, in order to conclude whether the current data vector contains outlier noise or not. Furthermore, a sparsity-promoting greedy algorithm is employed for the outlier vector estimation allowing the purification of the corrupted data from the outlier noise prior further processing. A theoretical analysis is carried out and experiments within the context of robust subspace estimation exhibit the enhanced performance of the proposed scheme compared to a recently developed state of the art algorithm.

In this paper, a distributed adaptive algorithm for sparsity-aware learning in diffusion networks is developed. The algorithm follows the greedy roadmap for sparsity along with the adapt-combine co-operation strategy, based on the LMS rationale for adaptivity. A bound on the error norm between the obtained estimates and the target vector is computed, and the algorithm is shown to converge in the mean under some general assumptions. Finally, comparative experiments with a recently developed sparsity-promoting diffusion LMS demonstrate the enhanced performance of the proposed algorithm.

This paper deals with a novel class of set-theoretic adaptive sparsity promoting algorithms of linear computational complexity. Sparsity is induced via generalized thresholding operators, which correspond to nonconvex penalties such as those used in a number of sparse LMS based schemes. The results demonstrate the significant performance gain of our approach, at comparable computational cost.

This letter establishes a novel analysis of the Adaptive Projected Subgradient Method (APSM) in the intersection of the stochastic and robust estimation paradigms. Utilizing classical worst-case bounds on the noise process, drawn from the robust estimation methodology, the present study demonstrates that the hyperslab-inspired version of the APSM generates a sequence of estimates which converges to a point located, with probability one, arbitrarily close to the estimand. Numerical tests and comparisons with classical time-adaptive algorithms corroborate the theoretical findings of the study.

In this paper, a novel scheme for online, sparsity-aware learning is presented. A new theory is developed that allows for the incorporation, in a unifying way, of different thresholding rules to promote sparsity, that may even be of a nonconvex nature. The complexity of the algorithm exhibits a linear dependence on the number of free parameters.

This paper presents a wide framework for non-linear online supervised learning tasks in the context of complex valued signal processing. The (complex) input data are mapped into a complex reproducing kernel Hilbert space (RKHS), where the learning phase is taking place. Both pure complex kernels and real kernels (via the complexification trick) can be employed. Moreover, any convex, continuous and not necessarily differentiable function can be used to measure the loss between the output of the specific system and the desired response. The only requirement is the subgradient of the adopted loss function to be available in an analytic form. In order to derive analytically the subgradients, the principles of the (recently developed) Wirtinger's calculus in complex RKHS are exploited. Furthermore, both linear and widely linear (in RKHS) estimation filters are considered. To cope with the problem of increasing memory requirements, which is present in almost all online schemes in RKHS, the sparsification scheme, based on projection onto closed balls, has been adopted. We demonstrate the effectiveness of the proposed framework in a non-linear channel identification task, a non-linear channel equalization problem and a quadrature phase shift keying equalization scheme, using both circular and non circular synthetic signal sources.

This paper introduces a wide framework for online, i.e., time-adaptive, supervised multiregression tasks. The problem is formulated in a general infinite-dimensional reproducing kernel Hilbert space (RKHS). In this context, a fairly large number of nonlinear multiregression models fall as special cases, including the linear case. Any convex, continuous, and not necessarily differentiable function can be used as a loss function in order to quantify the disagreement between the output of the system and the desired response. The only requirement is the subgradient of the adopted loss function to be available in an analytic form. To this end, we demonstrate a way to calculate the subgradients of robust loss functions, suitable for the multiregression task. As it is by now well documented, when dealing with online schemes in RKHS, the memory keeps increasing with each iteration step. To attack this problem, a simple sparsification strategy is utilized, which leads to an algorithmic scheme of linear complexity with respect to the number of unknown parameters. A convergence analysis of the technique, based on arguments of convex analysis, is also provided. To demonstrate the capacity of the proposed method, the multiregressor is applied to the multiaccess multiple-input multiple-output channel equalization task for a setting with poor resources and nonavailable channel information. Numerical results verify the potential of the method, when its performance is compared with those of the state-of-the-art linear techniques, which, in contrast, use space-time coding, more antenna elements, as well as full channel information.

In this paper, a novel algorithm for bandwidth reduction in adaptive distributed learning is introduced. We deal with diffusion net works, in which the nodes cooperate with each other, by exchanging information, in order to estimate an unknown parameter vector of interest. We seek for solutions in the framework of set theoretic estimation. Moreover, in order to reduce the required bandwidth by the transmitted information, which is dictated by the dimension of the unknown vector, we choose to project and work in a lower dimension Krylov subspace. This provides the benefit of trading off dimensionality with accuracy. Full convergence properties are presented, and experiments, within the system identification task, demonstrate the robustness of the algorithmic technique.

This paper presents a novel time-adaptive estimation technique by revisiting the classical Wiener-Hopf equation. Any convex and not necessarily differentiable function can be used for enlarging the Wiener-Hopf equation in order to incorporate the often met, in practice, measurement and model inaccuracies. Unlike classical techniques, e.g., the Recursive Least Squares (RLS) algorithm, the proposed method is free of the computation of the inverse of a correlation matrix. Moreover, the method offers the means for dealing with the presence of convex constraints in an efficient way, by exploiting general convex analytic tools. To validate the pro posed estimation method, an application of increasing importance nowadays, the online sparse signal recovery task is considered. Numerical results support the introduced theoretical arguments against the sparsity-aware classical batch, and the very recently introduced RLS-based signal recovery techniques.

This paper presents a novel projection-based adaptive algorithm for sparse system identification. Sequentially observed data are used to generate an equivalent number of closed convex sets, namely hyperslabs, which quantify an associated cost criterion. Sparsity is exploited by the introduction of appropriately designed weighted ℓ1balls. The algorithm uses only projections onto hyperslabs and weighted ℓ1balls, and results into a computational complexity of order O(L) multiplications/additions and O(Llog2L) sorting operations, where L is the length of the system to be estimated. Numerical results are also given to validate the proposed method against very recently developed sparse LMS and RLS type of algorithms, which are considered to belong to the same type of algorithmic family.

The goal of this paper is the development of a novel approach for the problem of Noise Removal, based on the theory of Reproducing Kernels Hilbert Spaces (RKHS). The problem is cast as an optimization task in a RKHS, by taking advantage of the celebrated semi parametric Representer Theorem. Examples verify that in the presence of gaussian noise the proposed method performs relatively well compared to wavelet based techniques and outperforms them significantly in the presence of impulse or mixed noise.

This paper presents a method for detecting violent content in video sharing sites. The proposed approach operates on a fusion of three modalities: audio, moving image and text data, the latter being collected from the accompanying user comments. The problem is treated as a binary classification task (violent vs non-violent content) on a 9-dimensional feature space, where 7 out of 9 features are extracted from the audio stream. The proposed method has been evaluated on 210 YouTube videos and the overall accuracy has reached 82%.

The main contribution of this paper is the development of a novel approach, based on the theory of Reproducing Kernel Hilbert Spaces (RKHS), for the problem of noise removal in the spatial domain. The proposed methodology has the advantage that it is able to remove any kind of additive noise (impulse, gaussian, uniform, etc.) from any digital image, in contrast to the most commonly used denoising techniques, which are noise dependent. The problem is cast as an optimization task in a RKHS, by taking advantage of the celebrated Representer Theorem in its semi-parametric formulation. The semi-parametric formulation, although known in theory, has so far found limited, to our knowledge, application. However, in the image denoising problem, its use is dictated by the nature of the problem itself. The need for edge preservation naturally leads to such a modeling. Examples verify that in the presence of gaussian noise the proposed methodology performs well compared to wavelet based technics and outperforms them significantly in the presence of impulse or mixed noise.

In this paper we present a novel method for extracting affective information from movies, based on speech data. The method is based on a 2D representation of speech emotions (Emotion Wheel). The goal is twofold. First, to investigate whether the Emotion Wheel offers a good representation for emotions associated with speech signals. To this end, several humans have manually annotated speech data from movies using the Emotion Wheel and the level of disagreement has been computed as a measure of representation quality. The results indicate that the emotion wheel is a good representation of emotions in speech data. Second, a regression approach is adopted, in order to predict the location of an unknown speech segment in the Emotion Wheel. Each speech segment is represented by a vector of ten audio features. The results indicate that the resulting architecture can estimate emotion states of speech from movies, with sufficient accuracy.

This paper presents a novel method for incorporating a-priori affine constraints in online kernel-based learning tasks. The proposed technique elaborates the generic tool of projections to form a sequence of estimates in reproducing kernel Hilbert spaces (RKHS). The method guarantees that the whole sequence of estimates lies in the given affine constraint set. To validate the algorithm, a beamforming task is considered. The numerical results show that the proposed frame provides with solutions in cases where the classical linear approach collapses, and forms proper beam-patterns as opposed to a recent unconstrained kernel-based regression method.

This paper treats gunshot detection in audio streams from movies as a maximization task, where the solution is obtained by means of dynamic programming. The proposed method seeks the sequence of segments and respective class labels, i.e., gunshots vs. all other audio types, that maximize the product of posterior class label probabilities, given the segments' data. The required posterior probabilities are estimated by combining soft classification decisions from a set of Bayesian Network combiners. Tests that have been performed on a large set of audio streams indicate that the proposed method yields high performance in terms of both precision and recall of detected gunshot events.

In this paper, a novel approach to audio segmentation is presented. The problem of detecting audio segmentspsila limits is treated as a binary classification task. Frames are classified as ldquosegment limitsrdquo vs ldquononsegment limitsrdquo. For each audio frame a spectrogram is computed and eight feature values are extracted from respective frequency bands. Final decisions are taken based on a classifier combination scheme. The algorithm has very low complexity with almost real time performance. It achieves 86% accuracy rate on real audio streams extracted from movies. Moreover, it introduces a general framework to audio segmentation, which does not depend explicitly on the number of audio classes.

Very recently, an adaptive projection algorithm was introduced for the online classification task with sparsification in reproducing kernel Hilbert spaces (RKHS). This paper presents another sparsification method for the projection-based approach by generating a sequence of linear subspaces in RKHS. Projection mappings give a geometrical flavor to the design; classification is performed by metric projection mappings, sparsification is achieved by orthogonal projections, while the online system's memory and tracking requirements are attained by oblique projections. The resulting sparsification scheme shows strong similarities with the classical sliding window adaptive schemes. Validation is performed by considering the adaptive equalization problem of a nonlinear communication channel. Although here the classification scheme is considered, the method is readily extended to regression tasks. Furthermore its generality allows for a number of cost functions including non-differentiable ones.

This paper presents a robust and computationally efficient method for tracking music in audio streams from movies. The audio stream is first mid-term processed with a fixed length moving window and four features are extracted per window. Each feature is fed as input to a simple classifier which produces a soft output for the binary problem of music vs. all other types of audio. The soft outputs are then combined to yield a measure of confidence quantifying whether the segment corresponds to music or not. At a final step, thresholding is applied to filter out segments where the confidence measure is low. The proposed approach has been tested with audio streams from various movies and its performance was measured both on a mid-term segment basis as well as on an event detection basis. Reported results demonstrate that the method exhibits high performance even when music is mixed with other types of audio in the stream.

This paper presents a multistage system for speech/music discrimination which is based on a three-step procedure. The first step is a computationally efficient scheme consisting of a region growing technique and operates on a 1-D feature sequence, which is extracted from the raw audio stream. This scheme is used as a preprocessing stage and yields segments with high music and speech precision at the expense of leaving certain parts of the audio recording unclassified. The unclassified parts of the audio stream are then fed as input to a more computationally demanding scheme. The latter treats speech/music discrimination of radio recordings as a probabilistic segmentation task, where the solution is obtained by means of dynamic programming. The proposed scheme seeks the sequence of segments and respective class labels (i.e., speech/music) that maximize the product of posterior class probabilities, given the data that form the segments. To this end, a Bayesian Network combiner is embedded as a posterior probability estimator. At a final stage, an algorithm that performs boundary correction is applied to remove possible errors at the boundaries of the segments (speech or music) that have been previously generated. The proposed system has been tested on radio recordings from various sources. The overall system accuracy is approximately 96%. Performance results are also reported on a musical genre basis and a comparison with existing methods is given.

This book provides a comprehensive and self-contained introduction to the field of pattern recognition (PR).

This work addresses the problem of locating rhythmic patterns in music recordings. During the feature extraction stage, a short-term processing technique is applied, in order to detect significant changes in the spectral and energy evolution of the music signal. The detected changes are in turn treated as onsets of events and a sequence of inter-onset intervals is extracted. The resulting sequence is long-term segmented and is fed as input to a hidden Markov model (HMM) which models a predefined rhythmic pattern. An enhanced Viterbi algorithm is proposed, that extracts a best-state sequence, which determines the pattern location boundaries. Our method was tested on a set of music recordings of music meter 2/4, 3/4, 7/8 and 9/8 and steady tempo. The proposed method exhibits excellent precision (100%) over pattern locations and a recall ranging from ~ 34% up to ~ 74% depending on the music genre.

The goal of this paper is the development of a novel efficient online kernel-based algorithm for classification. The spirit of the algorithm stems from the recently introduced adaptive projected subgradient method. This is a general convex analytic tool that employs projections onto a sequence of convex sets and it can be considered as a generalization of the celebrated APA algorithm, widely used in classical adaptive filtering.

A new method for object tracking in video sequences is presented. This method exploits the benefits of particle filters to tackle the multimodal distributions emerging from cluttered scenes. The tracked object is described by several models of different complexity, which are probabilistically linked together. The parameter update for each model takes place hierarchically so that the simpler models, which are updated first, can guide the search in the parameter space of the more complex models to relevant regions. This strategy improves the target representation because of the multiple models and reduces the overall complexity. The likelihood for each object model is calculated using one or more visual cues thus increasing the robustness of the proposed algorithm. Our method is evaluated by fusing on salient points and contour models and we demonstrate its effectiveness.

In this paper, training-based estimation of correlated block fading channels in a multiple-antenna, multi-user environment is considered. The linear minimum mean squared error (LMMSE) estimator is presented first. Then the problem of optimally designing the training data set, so as to minimize the mean squared channel estimation error subject to a total transmit power constraint, is addressed. It is shown that the optimal transmission directions are dictated jointly by the eigen-decompositions of the channel and interference covariance matrices. Their roles, in the channel estimation and interference suppression tasks, respectively, are revealed in the optimal transmit beamformer structure. The simulation results demonstrate that the gain in estimation performance from using the optimal training sequence increases considerably with increasing spatial fading correlation, especially in strong interference environments.

In this work, we present a multi-class classification algorithm for audio segments recorded from movies, focusing on the detection of violent content, for protecting sensitive social groups (e.g. children). Towards this end, we have used twelve audio features stemming from the nature of the signals under study. In order to classify the audio segments into six classes (three of them violent), Bayesian networks have been used in combination with the one versus all classification architecture. The overall system has been trained and tested on a large data set (5000 audio segments), recorded from more than 30 movies of several genres. Experiments showed, that the proposed method can be used as an accurate multi-class classification scheme, but also, as a binary classifier for the problem of violent -non violent audio content.

Geometric methods are very intuitive and provide a theoretically solid approach to many optimization problems. One such optimization task is the support vector machine (SVM) classification, which has been the focus of intense theoretical as well as application-oriented research in machine learning. In this letter, the incorporation of recent results in reduced convex hulls (RCHs) to a nearest point algorithm (NPA) leads to an elegant and efficient solution to the SVM classification task, with encouraging practical results to real-world classification problems, i.e., linear or nonlinear and separable or nonseparable.

This paper presents a speech/music discriminator for radio recordings. The segmentation stage is based on the detection of changes in the energy distribution of the audio signal. For the classification stage, Bayesian networks have been adopted in order to combine the results of nine k-nearest neighbor classifiers trained on individual features. To this end, a comparison of the performance of three popular Bayesian network architectures is presented. Furthermore, in order to reduce the number of features used for classification, a new feature selection scheme is introduced, that is also based on the properties of Bayesian networks. The proposed system has been tested on real Internet broadcasts of BBC radio stations

This paper presents a new extension to the variable duration hidden Markov model (HMM), capable of classifying musical pattens that have been extracted from raw audio data into a set of predefined classes. Each musical pattern is converted into a sequence of music intervals by means of a fundamental frequency tracking procedure. This sequence is subsequently presented as input to a set of variable-duration HMMs. Each one of these models has been trained to recognize patterns of a corresponding predefined class. Classification is determined based on the highest recognition probability. The new type of variable-duration hidden Markov modeling proposed in this paper results in enhanced performance because 1) it deals effectively with errors that commonly originate during the feature extraction stage, and 2) it accounts for variations due to the individual expressive performance of different instrument players. To demonstrate its effectiveness, the novel classification scheme has been employed in the context of Greek traditional music, to monophonic musical patterns of a popular instrument, the Greek traditional clarinet. Although the method is also appropriate for western-style music, Greek traditional music poses extra difficulties and makes music pattern recognition a harder task. The classification results demonstrate that the new approach outperforms previous work based on conventional HMMs

The geometric framework for the support vector machine (SVM) classification problem provides an intuitive ground for the understanding and the application of geometric optimization algorithms, leading to practical solutions of real world classification problems. In this work, the notion of "reduced convex hull" is employed and supported by a set of new theoretical results. These results allow existing geometric algorithms to be directly and practically applied to solve not only separable, but also nonseparable classification problems both accurately and efficiently. As a practical application of the new theoretical results, a known geometric algorithm has been employed and transformed accordingly to solve nonseparable problems successfully.

Recently, a novel MLSE equalizer was reported, that does not require the explicit estimation of the channel impulse response. Instead, it utilizes, in an efficient manner, the estimates of the centers of the clusters formed by the received observations. In this paper, a novel cluster tracking scheme is presented, which extends the application of this equalizer in time-varying transmission environments. The proposed algorithm is shown to be equivalent in tracking performance with the classic LMS-based MLSE equalizer, yet much simpler computationally. This is a consequence of the fact that the new method allows for an efficient exploitation of the symmetries underlying the signaling scheme.

Automatic recognition of musical patterns plays a crucial part in musicological and ethnomusicological research and can become an indispensable tool for the search and comparison of music extracts within a large multimedia database. This paper presents an efficient method for recognizing isolated musical patterns in a monophonic environment, using a novel extension of dynamic time warping, which we call context dependent dynamic time warping. Each pattern, to be recognized, is converted into a sequence of frequency jumps by means of a fundamental frequency tracking algorithm, followed by a quantizer. The resulting sequence of frequency jumps is presented to the input of the recognizer. The main characteristic of context dependent dynamic time warping is that it exploits the correlation exhibited among adjacent frequency jumps of the feature sequence. The methodology has been tested in the context of Greek traditional music, which exhibits certain characteristics that make the classification task harder, when compared with western musical tradition. A recognition rate higher than 95% was achieved.

We present the development of a system for the off-line optical recognition of the characters of the Orthodox Hellenic Byzantine Music Notation, that has been established since 1814. We describe the structure of the system, and propose algorithms for the recognition of the 71 distinct character classes, based on wavelets, 4-projections and other structural and statistical features. Using a simple nearest neighbor classifier and a tree-structured classification schema, an accuracy of 99.3% was achieved, in a database of about 18000 Byzantine music character patterns.

A new block adaptive decision feedback equalizer is developed. Both the feedforward (FF) and the feedback (FB) filters are updated once every K sample time intervals, with K being the block length. It should be noted that this block adaptation is done in such a way that the resulting filters, and the decisions as well, are identical to those computed by the conventional sample by sample LMS-based decision feedback equalizer (LMS-DFE). The new algorithm offers substantial computational savings as compared to the sample-by-sample LMS-DFE with no loss in performance. The new block DFE turns out to be particularly suitable for applications requiring long equalizers.

Transform-domain adaptive signal processing proved to be very successful in very many applications especially where systems with long impulse responses are to be evaluated. The popularity of these methods is due to the efficiency of the fast signal transformation algorithms and that of the block oriented adaptation mechanisms. The applicability of the fast sliding transformation algorithms is investigated for transform domain adaptive signal processing. It is shown that these sliding transformers may contribute to a better distribution of the computational load along time and therefore enable higher sampling rates. It is also shown that the execution time of the widely used overlap-save and overlap-add algorithms can also be shortened. The prize to be paid for this improvements is the increase of the end-to-end delay which in certain configurations may cause some degradation of the tracking capabilities of the overall system. Fortunately, however, there are versions where this delay does not hurt the capabilities of the adaptation technique applied.

The algorithm presented in the paper is an exact block processing counterpart of the fast Newton transversal filtering (FNTF) algorithm [Moustakides and Theodorides, 1991]. The main trait of the new algorithm is that the block processing is done in such a way so that the resulting estimates are mathematically equivalent with the respective estimates of the FNTF algorithm. In cases where the involved filter is of medium to long order the new algorithm offers a substantial saving in computational complexity without sacrificing performance.

The paper presents two new, closely related adaptive algorithms for LS system identification. The starting point for the derivation of the algorithms is the inverse Cholesky factor of the data correlation matrix, obtained via a QR decomposition (QRD). Both are of O(p) computational complexity with p being the order of the system. The first algorithm is a fixed order QRD scheme with enhanced parallelism. The second is a lattice type algorithm based on Givens rotations, with lower complexity compared to previously derived ones.

Fast recursive least squares (RLS) adaptive filters are still too complex for many applications. The authors discuss the fast Newton transversal filters (FNTFs), a new family of fast RLS filters based on prediction order reduction, which can have complexity close to that of LMS. A simple derivation of the FNTF is presented, and implementation, initialization, and stabilization issues, which are taken from fast RLS practice, are considered. When evaluating complexity, it is shown that FNTF filters with lengths appropriate for acoustic echo cancellation can be implemented on one standard floating point digital signal processor (DSP). Experimental performances on speechlike and real signals are very close to those of the standard fast RLS.>

During the past decade, systolic arrays have been designed for a wide variety of scientific applications, which are based on highly parallel linear system manipulations. Partitioning and mapping of systolic algorithms has been a key issue for real implementations, in terms of both cost and manageability. The authors demonstrate the mapping of triangular systolic array algorithms onto a one-dimensional ring of processors, so that the resulting architecture features an asymptotically optimal utilization factor in pipelined operation. fee problems of least squares system identification and FIR filtering using QR-decomposition via Givens rotations are used as a vehicle for the demonstration of uni- and bi-directional dataflow algorithms on systolic rings.>

The authors present a novel Levinson-type order recursive algorithm for the solution of block rho -Toeplitz systems of equations. Its main advantage is that it avoids matrix inversions. Thus, it can be used as the starting point for the derivation of the stairwise Schur-type algorithm which inherits high parallelism and, since it avoids matrix inversions, is suitable for VLSI implementation.>

A very powerful technique for computing the LS (least squares) estimates of an FIR (finite impulse response) filter's impulse response is described. It is based on the QR factorization of the input data matrix. The method consists of two parts. First the input matrix is factorized into an orthogonal Q part and an upper triangular R part. The unknown coefficients are then obtained from a triangular linear system of equations. An algorithm for solving the above linear system, which is appropriate for adaptive processing, is proposed. This is achieved via a set of Givens rotations and a modified Faddeeva scheme.>

The effects of a narrowband interference present in a pseudonoise (PN) spread-spectrum system can be minimized by using digital whitening techniques. A new efficient block LS algorithm for the design of an FIR filter with linear phase is derived and used as a whitening filter. Simulations are carried out to demonstrate the effectiveness of the LS optimum linear-phase filter in suppressing a narrowband interference in a PN spread-spectrum system. Comparisons to previously used methods are made. The simulations showed an improvement in the output SNR on the order of 4-5 dB over already existing schemes.>

Highly parallel algorithms are derived for multichannel FIR (finite-impulse response) smoothing and MEM (minimum-energy method) spectral analysis. The derived algorithms require O(p) computing time and can be performed on a linear array of O(p) processors, p being the order of the filter of AR (autoregressive) model. Thus, a computational saving of one order of magnitude is achieved, compared to Levinson-type algorithms.>

The effects of a narrowband interference in a PN spread spectrum system can be minimized by employing a whitening filter. The performance of this filter can be improved if its impulse response is symmetrical. In this paper an optimum LS FIR smoother is adopted to perform the whitening process and a new efficient algorithm is presented to compute the smoothers coefficients. To demonstrate the effectiveness of the LS FIR smoother in suppressing a narrowband interference, simulations are carried out and the results are compared with those obtained by previously employed techniques.