In this work, we introduce a novel methodology called Fusion Grad-CAM for generating attention maps in ensemble classification systems that utilize majority voting. Currently, there is no established technique for producing a consolidated attention map in such systems. Traditional approaches require inspecting individual attention maps from each contributing model, which is inefficient. Our proposed method integrates Grad-CAM produced attention maps by averaging or weighted averaging them based on both the predicted class and the confidence level of each model’s decision. This approach offers a holistic representation of the ensemble’s decision-making process. We validate our methodology using various combinations of well-known convolutional neural networks (CNNs) pretrained on ImageNet, demonstrating the effectiveness and clarity of the resulting attention maps.

In this study, we aim to predict health outcomes concerning dementia by leveraging data from individuals' first two doctor visits. Using the OASIS-2 dataset, we focus on early-stage, longitudinal data, which is especially crucial for patients initially diagnosed as healthy. This approach allows us to anticipate their future health trajectories more accurately. We propose and evaluate multiple machine learning models, finding the Extreme Gradient Boosting (XGBoost) algorithm to be particularly effective, with an accuracy rate of 100%. Our methodology provides a valuable resource for early interventions and preventive measures in dementia care.

In this paper, we propose an energy-efficient radar beampattern design framework for Millimeter Wave (mmWave) massive multi-input multi-output (mMIMO) systems, equipped with a hybrid analog-digital (HAD) beamforming structure. Aiming to reduce the power consumption and hardware cost of the mMIMO system, we employ a learning approach to synthesize the probing beampattern based on a small number of RF chains and antennas. By leveraging a combination of softmax neural networks, the proposed solution is able to achieve a desirable beampattern with high accuracy while incurring low cost.

Wireless Autonomous Networks are expected to support communication between a source and a receiver, by constantly self-adapting to changes in their communication environment. This paper considers a scenario of relay beamforming, in which relays collaboratively retransmit the source signal so that they maximize the average signal-to-interference+noise ratio (SINR) at the destination. The relays are grouped into clusters, with each cluster having a single active relay at a time. The system evolves in time slots; in each time slot the clusters beamform to the destination, and at the same time, each cluster selects the relay to be active in the subsequent time slot. Relay selection is performed locally within each cluster, using a reinforcement learning approach, namely Q-learning. Compared to prior methods, the proposed scheme does not require any statistical knowledge on the channels, and achieves similar average SINR performance while involving lower complexity.

The growth rates of today’s societies and the rapid advances in technology have led to the need for access to dynamic, adaptive and personalized information in real time. Augmented reality provides prompt access to rapidly flowing information which becomes meaningful and “alive” as it is embedded in the appropriate spatial and time framework. Augmented reality provides new ways for users to interact with both the physical and digital world in real time. Furthermore, the digitization of everyday life has led to an exponential increase of data volume and consequently, not only have new requirements and challenges been created but also new opportunities and potentials have arisen. Knowledge graphs and semantic web technologies exploit the data increase and web content representation to provide semantically interconnected and interrelated information, while deep learning technology offers novel solutions and applications in various domains. The aim of this study is to present how augmented reality functions and services can be enhanced when integrating deep learning, semantic web and knowledge graphs and to showcase the potentials their combination can provide in developing contemporary, user-friendly and user-centered intelligent applications. Particularly, we briefly describe the concept of augmented reality and mixed reality and present deep learning, semantic web and knowledge graphs technologies. Moreover, based on our literature review, we present and analyze related studies regarding the development of augmented reality applications and systems that utilize these technologies. Finally, after discussing how the integration of deep learning, semantic web and knowledge graphs into augmented reality enhances the quality of experience and quality of service of augmented reality applications to facilitate and improve users’ everyday life, conclusions and suggestions for future research and studies are given.

Sleep stage classification is one of the most critical steps in the effective diagnosis and treatment of sleep-related disorders. Classic approaches involve trained human sleep scorers, utilizing a manual scoring technique, according to certain standards. This paper examines the implementation of an algorithm for the automation of the sleep scoring process. EEG recordings data are acquired from three different groups comprising of healthy subjects and people with minor sleep disorders. A mixture of time domain and frequency domain features are extracted. Temporal feature changes are utilized in order to capture contextual information of the EEG signal. Multiple classifiers are tested, culminating in a voting classifier, achieving a maximum accuracy of 90.8% for the healthy subjects' group. The main novelty introduced by the proposed solution is the algorithm's high accuracy when tested on a mixed dataset of healthy and patient subjects. The promising capabilities that derive from the successful implementation of this solution are discussed in the conclusions.

This paper proposes a machine learning approach to part-of-speech tagging and named entity recognition for Greek, focusing on the extraction of morphological features and classification of tokens into a small set of classes for named entities. The architecture model that was used is introduced. The greek version of the spaCy platform was added into the source code, a feature that did not exist before our contribution, and was used for building the models. Additionally, a part of speech tagger was trained that can detect the morphology of the tokens and performs higher than the state-of-the-art results when classifying only the part of speech. For named entity recognition using spaCy, a model that extends the standard ENAMEX type (organization, location, person) was built. Certain experiments that were conducted indicate the need for flexibility in out-of-vocabulary words and there is an effort for resolving this issue. Finally, the evaluation results are discussed.

Deep convolutional neural networks have become the state of the art methods for image classification after demonstrating very good performance on very large datasets with general visual content. Amongst the problems for training deep CNN architectures is the heavy computational cost and the large memory requirements. In this work we exploit the fact that many training samples are correctly classified in the early stages of learning and therefore they can be skipped in further training epochs without noticeable performance degradation. We employ sample filtering to determine which samples are used in each epoch. We introduce the “multi-class margin” index to measure how safely a sample is classified to the correct class. The multi-class margin is closely connected with the output of the final Softmax layer. Using filtering we gradually reduce the number of usable samples to 40% of the original dataset. At the same time we gain around 18% in time compared to using all the samples in every epoch, while the performance degradation is around 1%.

In kernel-based classification models, given limited computational power and storage capacity, operations over the full kernel matrix becomes prohibitive. In this paper, we propose a new supervised learning framework using kernel models for sequential data processing. The framework is based on two components that both aim at enhancing the classification capability with a subset selection scheme. The first part is a subspace projection technique in the reproducing kernel Hilbert space using a CLAss-specific Subspace Kernel representation for kernel approximation. In the second part, we propose a novel structural risk minimization algorithm called the adaptive margin slack minimization to iteratively improve the classification accuracy by an adaptive data selection. We motivate each part separately, and then integrate them into learning frameworks for large scale data. We propose two such frameworks: the memory efficient sequential processing for sequential data processing and the parallelized sequential processing for distributed computing with sequential data acquisition. We test our methods on several benchmark data sets and compared with the state-of-the-art techniques to verify the validity of the proposed techniques.

In this paper, we design a novel regularized empirical risk minimization technique for classification called Adaptive Margin Slack Minimization (AMSM). The proposed method is based on minimizing a regularized upper bound of the misclassification error. Compared to the cost function of the classical L2-SVM, AMSM can be interpreted as minimizing a tighter bound with some additional flexibilities regarding the choice of marginal hyperplane. A hyperparameter-free adaptive algorithm is presented for finding a solution to the proposed risk function. Numerical results shows that AMSM outperforms L2-SVM on the tested standard datasets.

Signals in various applications are often generated by linear combinations of quantized components. The analysis of data into such components is treated here as a matrix analysis problem. We first show that the component alphabet can always be normalized to the levels 0, …, M-1, without loss of generality. Then we study certain conditions under which the decomposition is possible. In particular, we present an analytical algorithm based on the differences of the observed points and the recursive estimation of the quantized components when the number of unique observed points is sufficiently large.

The paper presents a new binary classification method based on the minimization of the slack variables energy called the Mean Squared Slack (MSS). We deliver preliminary mathematical results which support the motivation behind our approach. We show that (a) in the linearly separable case the minimum MSS is attained at a separating vector, while (b) the minimizer in the linearly non-separable case is bounded but not zero. The method is conceptually simple: it solves a linear system at each iteration and it converges, typically, within a few iterations. Its complexity is obviously related to the size of the system which, in the linear case, is equal to the input pattern dimension. The method is extended to the non-linear case using kernels. Simulations demonstrate that the method is competitive with respect to computation time, accuracy, and generalization performance compared to state of the art SVM methods.

We propose a new method for the blind separation of multiple binary signals from a single general nonlinear mixture. In addition to the usual independence assumption on the input signals our key hypothesis is the asymmetry of the source probabilities. This condition allows us to express the output probability distribution as a linear mixture of the sources. We then proceed to solve the problem using known linear BSS methods for the binary underdetermined case. The method is based on clustering avoiding costly iterative optimization. Our simulations demonstrate successful separation for up to four sources. The problem however grows exponentially with the number of sources n, and the dataset size required for accurate estimation may become prohibitively large for large n.

In this paper we present a novel blind method for separating two binary sources from a single, arbitrary nonlinear mixture. The method is analytical and does not involve nonlinear optimization. Our approach proceeds by linearizing the problem and extending known, clustering-based results from the linear binary BSS case to the nonlinear case. The proposed algorithm is computationally efficient. Due to the structure of the problem, the true sources are extracted together with a source product adding one more indeterminacy to the usual sign and order indeterminacy of the sources. In some applications (eg. imaging) this indeterminacy can be resolved by visual inspection.

We present a novel method for the blind identification problem of multichannel FIR systems based on the analytical solution of a linear system of equations derived from the output signal second order statistics. The method is very efficient computationally since it does not involve iterative optimization procedures. In addition, it is very robust to noise as shown by our experiments. We also demonstrate that it compares favorably in terms of estimation accuracy compared to classical second order methods.

We present two novel methods for the blind extraction of multiple binary signals from a single observation generated from linear convolution and superposition. Both methods are based on two features: (a) the constellation structure of the successor set for each output value; (b) the concept of system deflation, i.e. the recursive shortening of the channel until it is reduced to an instantaneous mixture. The two methods differ in complexity as the latter requires significantly smaller amounts of data. Both methods are fast for small problem sizes but their complexities increase exponentially as the number of inputs and/or channel length increases. We present the complete analytical solution to the noiseless case and we thoroughly treat the related mathematical tools and concepts. The noisy case is discussed briefly, as the complete treatment would exceed the scope of the paper.

A new method for the blind separation of linear image mixtures is presented in this paper. Such mixtures often occur, when, for example, we photograph a scene through a semireflecting medium (windshield or glass). The proposed method requires two mixtures of two scenes captured under different illumination conditions. We show that the boundary values of the ratio of the two mixtures can lead to an accurate estimation of the separation matrix. The technique is very simple, fast, and reliable, as it does not depend on iterative procedures. The method effectiveness is tested on both artificially mixed images and real images.

We solve the overdetermined MIMO blind deconvolution problem for independent, stationary, and temporally white sources. Our approach is based on the SVD analysis of the output autocorrelation matrices at different time lags. The key idea is the identification of suitable subspace projectors which achieve an effective reduction of the filter length. We refer to this process as "filter deflation". Recursive applications of the filter deflation transform simplify the problem into an instantaneous BSS problem.

A novel image sequence segmentation method which combines both spatial and temporal information is presented in this paper. The first step is an intensity segmentation scheme based on the edgeflow method. The temporal information is introduced through three-dimensional (3-D) motion estimation parameters. In the second step, regions obtained from the first step are clustered according to their 3-D motion models. In order to reduce the noise sensitivity of the motion estimation process, we introduce a robust method which produces accurate motion parameters and facilitates the correct clustering that follows. This ensures that rigid objects with luminance discontinuities can be segmented correctly. The method has been successfully tested in real imagery and typical examples are presented in this paper.

Oriented PCA (OPCA) is a (second order) extension of standard principal component analysis aiming at maximizing the power ratio of a pair of signals. It is shown that OPCA, preceded by almost arbitrary temporal filtering, can be used for blindly separating temporally colored signals from their linear instantaneous mixtures. The advantage over other second order techniques is the lack of the prewhitening (or sphereing) step. Although the design of the general optimal temporal pre-filter is an open problem, we show that the filters [1, /spl plusmn/1] are the optimal ones for the special two-tap case. Neural OPCA models proposed earlier are used in simulations to separate a number of artificial sources demonstrating the validity of the method.

We consider a general n × n MIMO system excited by unobservable inputs that are spatially independent, cyclostationary with unknown statistics. We provide a set of conditions under which the system is uniquely identifiable based on second-order frequency-domain correlations of the system output. Such a MIMO problem appears in many applications, such as multi-user communications and separation of competing speakers.

Object tracking with occlusion prediction using multiple feature correspondences is proposed. The tracking region is defined by a set of point features, tracked using Kanade-Lucas-Tomasi (1991) algorithm. During total occlusion the region position is estimated using motion prediction based on a Kalman filtering scheme applied to the motion model prior to occlusion. During partial occlusion the displacements of the occluded features are predicted based on the motion of the bounding box of the moving object. Experimental results on real and artificial images have shown that the algorithm behaves well under total and partial occlusion.

Multiple binary sequences are blindly separated from a single linear mixture using the structure of the probability distribution function of the observed data. No specific assumptions are made regarding the second or higher order statistics of the sources. Assuming additive Gaussian noise the PDF is a mixture of Gaussians centered at points that uniquely determine (a) the mixing parameters and (b) the source signals up to a permutation and a sign ambiguity. We present both the theoretical framework for this novel approach and a new recursive blind separation algorithm based on our framework. Simulations show that the method can successfully separate at least up to 10 binary source signals at different noise levels.

We present a new image sequence segmentation method which combines both spatial and temporal information in a multiresolution framework. A region growing technique in a multiresolution scheme outputs an over-segmented partition of the image scene. Pure temporal information is collected for each region using a feature extraction/feature tracking technique. Motion information is further processed in a pyramidal robust motion estimation scheme, in order to calculate more complex motion parameters for each region. Regions obtained from the first step are clustered according to their motion models. This ensures that rigid objects with luminance discontinuities can be segmented correctly. The method has been successfully tested in real imagery and typical examples are presented.

The estimation of rigid-body 3-D motion parameters using point correspondences from a pair of images under perspective projection is, typically, very sensitive to noise. We present a novel robust method combining two approaches: (1) the SVD analysis of a linear operator resulting from the feature points and the displacement vectors and (2) a modified version of the well-known weighted least-squares method proposed by Huber in the context of robust statistics. We give a detailed rank analysis of the involved linear operator and study the effects of noise. We also propose a robust method guided by the structure of this operator, using weighted least squares and data partitioning. The method has been tested on artificial data and on real image sequences showing a remarkable robustness, even in the presence of up to 50% outliers in the data set.

The blind estimation of mixing channels resulting from frequency selective fading and multipath in a multi-user CDMA system is an important problem in wireless communications. We present a novel frequency-domain approach using second order spectral statistics for recovering the unknown channels. Unlike other methods which are based on time-domain analysis we make no particular assumption about the support of the mixing channels except that they have finite length (FIR). The method is based on the fact that the source sequences obtain a known spectral color derived from the corresponding spreading code used in CDMA.

The estimation of rigid body 3-D motion parameters from perspective views is typically very sensitive to noise and also to the presence of outliers in the measurements. In this paper we present a robust 3-D motion estimation approach based on a previously proposed method using SVD analysis of the measurements matrix. On the introduction of noise and outliers the performance of the old method was seen to deteriorate rapidly. Here the problem is attached by splitting the measurement set into smaller subsets and combining the properties of the resulting submatrices with the properties of the desired solution vector in order to obtain our estimate. The method is very robust and it has been successfully tested in both artificial datasets and real images with up to 50% presence of outliers. In addition, the method is fast and more importantly, the estimate quality is independent of the percentage of outliers.

We treat the blind identification problem for a n/spl times/n MIMO system using second order frequency-domain statistics and asymmetric PCA neural models. It is assumed that the source signals are colored, stationary, and pair-wise independent sequences with otherwise unknown statistics. We introduce a set of invariant indices that are used to tackle the problem of frequency-dependent ambiguities in the ordering and the phase of the retrieved singular vectors. For the case of 2/spl times/2 systems we present a complete identification procedure based on the corresponding invariant indices and we conjecture that these indices can be instrumental in the solution of the general n/spl times/n problem as well.

The optimal linear Karhunen-Loeve transform (KLT) attains the minimum reconstruction error for a fixed number of transform coefficients assuming that these coefficients do not contain noise. In any real coding system, however, the representation of the coefficients using a finite number of bits requires the presence of quantizers. We formulate the optimal linear transform using a data model that incorporates the quantization noise. Our solution does not correspond to an orthogonal transform and in fact, it achieves a smaller mean squared error (MSE) compared to the KLT, in the noisy case. Like the KLT, our solution depends on the statistics of the input signal, but it also depends on the bit-rate used for each coefficient. Especially for images, based on our optimality theory, we propose a simple modification of the discrete cosine transform (DCT). Our coding experiments show a peak signal-to noise ratio (SNR) performance improvement over JPEG of the order of 0.2 dB with an overhead less than 0.01 b/pixel.

In a typical linear data compression system the representation variables resulting from the coding operation are assumed totally reliable and therefore the solution in the mean-squared-error sense is an orthogonal projector to the so-called principal component subspace. When the representation variables are contaminated by additive noise which is uncorrelated with the signal, the problem is called noisy principal component analysis (NPCA) and the optimal MSE solution is not a trivial extension of PCA. We first show that the problem is not well defined unless we impose explicit or implicit constraints on either the coding or the decoding operator. Second, orthogonality is not a property of the optimal solution under most constraints. Third, the signal components may or may not be reconstructed depending on the noise level. As the noise power increases, we observe rank reduction in the optimal solution under most reasonable constraints. In these cases it appears that it is preferable to omit the smaller signal components rather than attempting to reconstruct them. This phenomenon has similarities with classical information theoretical results, notably the water-filling analogy, found in parallel additive Gaussian noise channels. Finally, we show that standard Hebbian-type PCA learning algorithms are not optimally robust to noise, and propose a new Hebbian-type learning algorithm which is optimally robust in the NPCA sense.

A new method for estimating 3D motion parameters from point correspondences is presented in this paper. The problem formulation leads to the solution of an overdetermined linear system of equations. The total least squares (TLS) method is found to be the most suitable one for estimating the solution since our model includes noise both in the observation data and in the system matrix. The translation parameters are obtained immediately from the above solution whereas the rotation parameters are estimated from the solution of another TLS problem. Tests of our method on artificial data and on real images show its robustness against Gaussian additive noise and against digitalization noise introduced by finite pixel resolution.

The linear threshold element (LTE), or perceptron, is a linear classifier with limited capabilities due to the problems arising when the input pattern set is linearly nonseparable. Assuming that the patterns are presented in a sequential fashion, we derive a theory for the detection of linear nonseparability as soon as it appears in the pattern set. This theory is based on the precise determination of the solution region in the weight space with the help of a special set of vectors. For this region, called the solution cone, we present a recursive computation procedure which allows immediate detection of nonseparability. The separability-violating patterns may be skipped so that, at the end, we derive a totally separable subset of the original pattern set along with its solution cone. The intriguing aspect of this algorithm is that it can be directly cast into a simple neural-network implementation. In this model the synaptic weights are committed (they are updated only once, and the only change that may happen after that is their destruction). This bears resemblance to the behavior of biological neural networks, and it is a feature unlike those of most other artificial neural techniques. Finally, by combining many such neural models we develop a learning procedure capable of separating convex classes.

Noisy principal component analysis (NPCA) was introduced previously as an extension of PCA in the assumption that the linear features are unreliable. The level of noise in the representation variables is found to have effects in the rank of the optimal solution resembling the water-filling analogy in information theory. The NPCA problem needs to be coupled with certain constraints so that it permits a finite solution. We present the solution of the NPCA problem under different constraints which can be useful in applications involving bandwidth limitations. One of these applications is the design of optimal subband coders incorporating quantization noise. In addition to the NPCA-optimality another advantage of the new design approach is that it works entirely in the time domain and thus the costly and difficult transformations to and from the Z-domain can be avoided.

We present a very simple solution of the motion recovery problem using the displacement map produced by a camera moving in a stationary environment under the perspective projection assumption. If the motion equations hold exactly (ideal case), the solution is derived immediately from the singular value decomposition (SVD) of a linear system matrix. In the ideal case the rank of the matrix indicates whether there is a zero or non-zero translation component in the camera motion. In practice the motion equations do not hold exactly. We speculate that the solution vector is a linear combination of the 3 smallest but non-zero eigenvalue eigenvectors. Then we find the 3 combination coefficients so that the solution obeys an equal number of problem constraints. The speculation turns out to yield good results for small noise situations but degrades fast as the noise power increases.

This paper is developed in two parts. First, the authors formulate the solution to the general reduced-rank linear approximation problem relaxing the invertibility assumption of the input autocorrelation matrix used by previous authors. The authors' treatment unifies linear regression, Wiener filtering, full rank approximation, auto-association networks, SVD and principal component analysis (PCA) as special cases. The authors' analysis also shows that two-layer linear neural networks with reduced number of hidden units, trained with the least-squares error criterion, produce weights that correspond to the generalized singular value decomposition of the input-teacher cross-correlation matrix and the input data matrix. As a corollary the linear two-layer backpropagation model with reduced hidden layer extracts an arbitrary linear combination of the generalized singular vector components. Second, the authors investigate artificial neural network models for the solution of the related generalized eigenvalue problem. By introducing and utilizing the extended concept of deflation (originally proposed for the standard eigenvalue problem) the authors are able to find that a sequential version of linear BP can extract the exact generalized eigenvector components. The advantage of this approach is that it's easier to update the model structure by adding one more unit or pruning one or more units when the application requires it. An alternative approach for extracting the exact components is to use a set of lateral connections among the hidden units trained in such a way as to enforce orthogonality among the upper- and lower-layer weights. The authors call this the lateral orthogonalization network (LON) and show via theoretical analysis-and verify via simulation-that the network extracts the desired components. The advantage of the LON-based model is that it can be applied in a parallel fashion so that the components are extracted concurrently. Finally, the authors show the application of their results to the solution of the identification problem of systems whose excitation has a non-invertible autocorrelation matrix. Previous identification methods usually rely on the invertibility assumption of the input autocorrelation, therefore they can not be applied to this case.

The concept of oriented principal component (OPC) analysis is introduced. It is the extension of the GSVD (generalized singular value decomposition) concept to the case of random processes (much like principal component analysis extends SVD for stochastic signals). In the random signal case, OPC analysis is equivalent to matched filtering and can be found useful in many classification and detection applications. The authors propose a corresponding neural model equipped with an efficient training algorithm for estimating the oriented principal component of two stochastic processes without assuming explicit knowledge of their statistics. The algorithm is based on the (normalized) learning rule proposed by Hebb for training the synaptic weights of a network of neurons. Both the theoretical justification and the numerical performance are shown, giving an explicit estimate of the learning rate parameter for best convergence speed.>

The authors propose a transition count method for detecting faults in single- and multiple-output logic circuits. It can be extended to sequential circuits in which scan design is incorporated. This method is called double transition count (DTC) testing for single-output circuits and multiple transition count (MTC) testing for multiple-output circuits. It is shown that the detectability of faults obtained by DTC and MTC testing is the same as that obtained by conventional testing. Hence, this method does not result in any information loss even though the set of output vectors is considerably compressed. The size of a DTC or MTC test is equal to the size of the equivalent conventional test since no test vectors need to be repeated. The basic test circuitry required is very simple, consisting of only one flip flop, one OR gate, one inverter, and one switch per output.>

The problem of the recursive computation of the principal components of a vector stochastic process is discussed. The applications of this problem arise in modeling of control systems, high-resolution spectrum analysis, image data compression, motion estimation, etc. An algorithm called APEX which can recursively compute the principal components using a linear neural network is proposed. The algorithm is recursive and adaptive: given the first m-1 principal components, it can produce the mth component iteratively. The numerical theoretical basis of the fast convergence of the APEX algorithm is given, and its computational advantages over previously proposed methods are demonstrated. Extension to extracting constrained principal components using APEX is also discussed.>