Search

Speaker recognition is a task of identifying persons from their voices. Recently, deep learning has dramatically revolutionized speaker recognition. However, there is lack of comprehensive reviews on the exciting progress. In this paper, we review several major subtasks of speaker recognition, including speaker verification, identification, diarization, and robust speaker recognition, with a focus on deep-learning-based methods. Because the major advantage of deep learning over conventional methods is its representation ability, which is able to produce highly abstract embedding features from utterances, we first pay close attention to deep-learning-based speaker feature extraction, including the inputs, network structures, temporal pooling strategies, and objective functions respectively, which are the fundamental components of many speaker recognition subtasks. Then, we make an overview of speaker diarization, with an emphasis of recent supervised, end-to-end, and online diarization. Finally, we survey robust speaker recognition from the perspectives of domain adaptation and speech enhancement, which are two major approaches of dealing with domain mismatch and noise problems. Popular and recently released corpora are listed at the end of the paper.

In theoretical analysis of deep learning, discovering which features of deep learning lead to good performance is an important task. In this paper, using the framework for analyzing the generalization error developed in Suzuki (2018), we derive a fast learning rate for deep neural networks with more general activation functions. In Suzuki (2018), assuming the scale invariance of activation functions, the tight generalization error bound of deep learning was derived. They mention that the scale invariance of the activation function is essential to derive tight error bounds. Whereas the rectified linear unit (ReLU; Nair and Hinton, 2010) satisfies the scale invariance, the other famous activation functions including the sigmoid and the hyperbolic tangent functions, and the exponential linear unit (ELU; Clevert et al., 2016) does not satisfy this condition. The existing analysis indicates a possibility that a deep learning with the non scale invariant activations may have a slower convergence rate of $O(1/\sqrt{n})$ when one with the scale invariant activations can reach a rate faster than $O(1/\sqrt{n})$. In this paper, without the scale invariance of activation functions, we derive the tight generalization error bound which is essentially the same as that of Suzuki (2018). From this result, at least in the framework of Suzuki (2018), it is shown that the scale invariance of the activation functions is not essential to get the fast rate of convergence. Simultaneously, it is also shown that the theoretical framework proposed by Suzuki (2018) can be widely applied for analysis of deep learning with general activation functions., Comment: 32 pages

Many recent studies on online seizure prediction from iEEG signal describe various prediction algorithms and their prediction performance. In contrast, this paper focuses on proper specification of system parameters, such as prediction period, prediction horizon and data-driven characterization of lead seizures. Whereas prediction performance clearly depends on these system parameters many researchers simply set the values of these parameters in an ad hoc manner. Our paper investigates the effect of these system parameters on online prediction performance, using both synthetic and real-life data sets. Therefore, meaningful comparison of methods/algorithms (for online seizure prediction) should consider proper specification of system parameters.

This paper proves an abstract theorem addressing in a unified manner two important problems in function approximation: avoiding curse of dimensionality and estimating the degree of approximation for out-of-sample extension in manifold learning. We consider an abstract (shallow) network that includes, for example, neural networks, radial basis function networks, and kernels on data defined manifolds used for function approximation in various settings. A deep network is obtained by a composition of the shallow networks according to a directed acyclic graph, representing the architecture of the deep network. In this paper, we prove dimension independent bounds for approximation by shallow networks in the very general setting of what we have called G -networks on a compact metric measure space, where the notion of dimension is defined in terms of the cardinality of maximal distinguishable sets, generalizing the notion of dimension of a cube or a manifold. Our techniques give bounds that improve without saturation with the smoothness of the kernel involved in an integral representation of the target function. In the context of manifold learning, our bounds provide estimates on the degree of approximation for an out-of-sample extension of the target function to the ambient space. One consequence of our theorem is that without the requirement of robust parameter selection, deep networks using a non-smooth activation function such as the ReLU, do not provide any significant advantage over shallow networks in terms of the degree of approximation alone.

In this paper, we address the stability of a broad class of discrete-time hypercomplex-valued Hopfield-type neural networks. To ensure the neural networks belonging to this class always settle down at a stationary state, we introduce novel hypercomplex number systems referred to as real-part associative hypercomplex number systems. Real-part associative hypercomplex number systems generalize the well-known Cayley-Dickson algebras and real Clifford algebras and include the systems of real numbers, complex numbers, dual numbers, hyperbolic numbers, quaternions, tessarines, and octonions as particular instances. Apart from the novel hypercomplex number systems, we introduce a family of hypercomplex-valued activation functions called $\mathcal{B}$-projection functions. Broadly speaking, a $\mathcal{B}$-projection function projects the activation potential onto the set of all possible states of a hypercomplex-valued neuron. Using the theory presented in this paper, we confirm the stability analysis of several discrete-time hypercomplex-valued Hopfield-type neural networks from the literature. Moreover, we introduce and provide the stability analysis of a general class of Hopfield-type neural networks on Cayley-Dickson algebras.

This paper is motivated by an open problem around deep networks, namely, the apparent absence of over-fitting despite large over-parametrization which allows perfect fitting of the training data. In this paper, we analyze this phenomenon in the case of regression problems when each unit evaluates a periodic activation function. We argue that the minimal expected value of the square loss is inappropriate to measure the generalization error in approximation of compositional functions in order to take full advantage of the compositional structure. Instead, we measure the generalization error in the sense of maximum loss, and sometimes, as a pointwise error. We give estimates on exactly how many parameters ensure both zero training error as well as a good generalization error. We prove that a solution of a regularization problem is guaranteed to yield a good training error as well as a good generalization error and estimate how much error to expect at which test data., Comment: 21 pages; Accepted for publication in Neural Networks

This paper presents a novel adaptive resonance theory (ART)-based modular architecture for unsupervised learning, namely the distributed dual vigilance fuzzy ART (DDVFA). DDVFA consists of a global ART system whose nodes are local fuzzy ART modules. It is equipped with distributed higher-order activation and match functions and a dual vigilance mechanism. Together, these allow DDVFA to perform unsupervised modularization, create multi-prototype cluster representations, retrieve arbitrarily-shaped clusters, and reduce category proliferation. Another important contribution is the reduction of order-dependence, an issue that affects any agglomerative clustering method. This paper demonstrates two approaches for mitigating order-dependence: pre-processing using visual assessment of cluster tendency (VAT) or post-processing using a novel Merge ART module. The former is suitable for batch processing, whereas the latter also works for online learning. Experimental results in online mode carried out on 30 benchmark data sets show that DDVFA cascaded with Merge ART statistically outperformed the best other ART-based systems when samples were randomly presented. Conversely, they were found to be statistically equivalent in offline mode when samples were pre-processed using VAT. Remarkably, performance comparisons to non-ART-based clustering algorithms show that DDVFA (which learns incrementally) was also statistically equivalent to the non-incremental (offline) methods of density-based spatial clustering of applications with noise (DBSCAN), single linkage hierarchical agglomerative clustering (SL-HAC), and k-means, while retaining the appealing properties of ART. Links to the source code and data are provided. Considering the algorithm’s simplicity, online learning capability, and performance, it is an ideal choice for many agglomerative clustering applications.

Building spiking neural networks (SNNs) based on biological synaptic plasticities holds a promising potential for accomplishing fast and energy-efficient computing, which is beneficial to mobile robotic applications. However, the implementations of SNNs in robotic fields are limited due to the lack of practical training methods. In this paper, we therefore introduce both indirect and direct end-to-end training methods of SNNs for a lane-keeping vehicle. First, we adopt a policy learned using the \textcolor{black}{Deep Q-Learning} (DQN) algorithm and then subsequently transfer it to an SNN using supervised learning. Second, we adopt the reward-modulated spike-timing-dependent plasticity (R-STDP) for training SNNs directly, since it combines the advantages of both reinforcement learning and the well-known spike-timing-dependent plasticity (STDP). We examine the proposed approaches in three scenarios in which a robot is controlled to keep within lane markings by using an event-based neuromorphic vision sensor. We further demonstrate the advantages of the R-STDP approach in terms of the lateral localization accuracy and training time steps by comparing them with other three algorithms presented in this paper.

This paper carries out a theoretical investigation into the stability problem for the class of neutral-type Cohen-Grossberg neural networks with discrete time delays in states and discrete neutral delays in time derivative of states. By employing a more general type of suitable Lyapunov functional, a set of new generalized sufficient criteria are derived for the global asymptotic stability of delayed neural networks of neutral-type. The proposed stability criteria are independently of the values of the time delays and neutral delays, and they completely rely on some algebraic mathematical relationships involving the values of the elements of the interconnection matrices and the other network parameters. Therefore, it is easy to verify the validity of the obtained results by simply using some algebraic equations representing the stability conditions. A detailed comparison between our proposed results and recently reported corresponding stability results is made, proving that the results given in this paper generalize previously published stability results. A constructive numerical example is also given to demonstrate the applicability of the results of the paper. (C) 2019 Elsevier Ltd. All rights reserved.

A recurrent neural network (RNN) can generate a sequence of patterns as the temporal evolution of the output vector. This paper focuses on a continuous-time RNN model with a piecewise-linear activation function that has neither external inputs nor hidden neurons, and studies the problem of finding the parameters of the model so that it generates a given sequence of bipolar vectors. First, a sufficient condition for the model to generate the desired sequence is derived, which is expressed as a system of linear inequalities in the parameters. Next, three approaches to finding solutions of the system of linear inequalities are proposed: One is formulated as a convex quadratic programming problem and others are linear programming problems. Then, two types of sequences of bipolar vectors that can be generated by the model are presented. Finally, the case where the model generates a periodic sequence of bipolar vectors is considered, and a sufficient condition for the trajectory of the state vector to converge to a limit cycle is provided.

Catastrophic forgetting (CF) happens whenever a neural network overwrites past knowledge while being trained on new tasks. Common techniques to handle CF include regularization of the weights (using, e.g., their importance on past tasks), and rehearsal strategies, where the network is constantly re-trained on past data. Generative models have also been applied for the latter, in order to have endless sources of data. In this paper, we propose a novel method that combines the strengths of regularization and generative-based rehearsal approaches. Our generative model consists of a normalizing flow (NF), a probabilistic and invertible neural network, trained on the internal embeddings of the network. By keeping a single NF throughout the training process, we show that our memory overhead remains constant. In addition, exploiting the invertibility of the NF, we propose a simple approach to regularize the network's embeddings with respect to past tasks. We show that our method performs favorably with respect to state-of-the-art approaches in the literature, with bounded computational power and memory overheads., Comment: arXiv admin note: substantial text overlap with arXiv:2007.02443

Spiking neural networks (SNNs) with event-based computation are promising brain-inspired models for energy-efficient applications on neuromorphic hardware. However, most supervised SNN training methods, such as conversion from artificial neural networks or direct training with surrogate gradients, require complex computation rather than spike-based operations of spiking neurons during training. In this paper, we study spike-based implicit differentiation on the equilibrium state (SPIDE) that extends the recently proposed training method, implicit differentiation on the equilibrium state (IDE), for supervised learning with purely spike-based computation, which demonstrates the potential for energy-efficient training of SNNs. Specifically, we introduce ternary spiking neuron couples and prove that implicit differentiation can be solved by spikes based on this design, so the whole training procedure, including both forward and backward passes, is made as event-driven spike computation, and weights are updated locally with two-stage average firing rates. Then we propose to modify the reset membrane potential to reduce the approximation error of spikes. With these key components, we can train SNNs with flexible structures in a small number of time steps and with firing sparsity during training, and the theoretical estimation of energy costs demonstrates the potential for high efficiency. Meanwhile, experiments show that even with these constraints, our trained models can still achieve competitive results on MNIST, CIFAR-10, CIFAR-100, and CIFAR10-DVS. Our code is available at https://github.com/pkuxmq/SPIDE-FSNN., Accepted by Neural Networks

This paper presents a computationally feasible method to compute rigorous bounds on the interval-generalisation of regression analysis to account for epistemic uncertainty in the output variables. The new iterative method uses machine learning algorithms to fit an imprecise regression model to data that consist of intervals rather than point values. The method is based on a single-layer interval neural network which can be trained to produce an interval prediction. It seeks parameters for the optimal model that minimizes the mean squared error between the actual and predicted interval values of the dependent variable using a first-order gradient-based optimization and interval analysis computations to model the measurement imprecision of the data. An additional extension to a multi-layer neural network is also presented. We consider the explanatory variables to be precise point values, but the measured dependent values are characterized by interval bounds without any probabilistic information. The proposed iterative method estimates the lower and upper bounds of the expectation region, which is an envelope of all possible precise regression lines obtained by ordinary regression analysis based on any configuration of real-valued points from the respective y-intervals and their x-values., Comment: 35 pages, 8 figures

Recently, deep Convolutional Neural Networks (CNNs) have proven to be successful when employed in areas such as reduced order modeling of parametrized PDEs. Despite their accuracy and efficiency, the approaches available in the literature still lack a rigorous justification on their mathematical foundations. Motivated by this fact, in this paper we derive rigorous error bounds for the approximation of nonlinear operators by means of CNN models. More precisely, we address the case in which an operator maps a finite dimensional input $\boldsymbol{\mu}\in\mathbb{R}^{p}$ onto a functional output $u_{\boldsymbol{\mu}}:[0,1]^{d}\to\mathbb{R}$, and a neural network model is used to approximate a discretized version of the input-to-output map. The resulting error estimates provide a clear interpretation of the hyperparameters defining the neural network architecture. All the proofs are constructive, and they ultimately reveal a deep connection between CNNs and the Fourier transform. Finally, we complement the derived error bounds by numerical experiments that illustrate their application.

The application of biologically inspired methods in design and control has a long tradition in robotics. Unlike previous approaches in this direction, the emerging field of neurorobotics not only mimics biological mechanisms at a relatively high level of abstraction but employs highly realistic simulations of actual biological nervous systems. Even today, carrying out these simulations efficiently at appropriate timescales is challenging. Neuromorphic chip designs specially tailored to this task therefore offer an interesting perspective for neurorobotics. Unlike Von Neumann CPUs, these chips cannot be simply programmed with a standard programming language. Like real brains, their functionality is determined by the structure of neural connectivity and synaptic efficacies. Enabling higher cognitive functions for neurorobotics consequently requires the application of neurobiological learning algorithms to adjust synaptic weights in a biologically plausible way. In this paper, we therefore investigate how to program neuromorphic chips by means of learning. First, we provide an overview over selected neuromorphic chip designs and analyze them in terms of neural computation, communication systems and software infrastructure. On the theoretical side, we review neurobiological learning techniques. Based on this overview, we then examine on-die implementations of these learning algorithms on the considered neuromorphic chips. A final discussion puts the findings of this work into context and highlights how neuromorphic hardware can potentially advance the field of autonomous robot systems. The paper thus gives an in-depth overview of neuromorphic implementations of basic mechanisms of synaptic plasticity which are required to realize advanced cognitive capabilities with spiking neural networks.

We present an extension of a neurobiologically inspired robotics model, termed CoRLEGO (Choice reaching with a LEGO arm robot). CoRLEGO models experimental evidence from choice reaching tasks (CRT). In a CRT participants are asked to rapidly reach and touch an item presented on the screen. These experiments show that non-target items can divert the reaching movement away from the ideal trajectory to the target item. This is seen as evidence attentional selection of reaching targets can leak into the motor system. Using competitive target selection and topological representations of motor parameters (dynamic neural fields) CoRLEGO is able to mimic this leakage effect. Furthermore if the reaching target is determined by its colour oddity (i.e. a green square among red squares or vice versa), the reaching trajectories become straighter with repetitions of the target colour (colour streaks). This colour priming effect can also be modelled with CoRLEGO. The paper also presents an extension of CoRLEGO. This extension mimics findings that transcranial direct current stimulation (tDCS) over the motor cortex modulates the colour priming effect (Woodgate et al., 2015). The results with the new CoRLEGO suggest that feedback connections from the motor system to the brain’s attentional system (parietal cortex) guide visual attention to extract movement-relevant information (i.e. colour) from visual stimuli. This paper adds to growing evidence that there is a close interaction between the motor system and the attention system. This evidence contradicts the traditional conceptualization of the motor system as the endpoint of a serial chain of processing stages. At the end of the paper we discuss CoRLEGO’s predictions and also lessons for neurobiologically inspired robotics emerging from this work.

In this paper, the synchronization problem for neural networks with stochastic perturbation is studied with intermittent control via adaptive aperiodicity. Under the framework of stochastic theory and Lyapunov stability method, we develop some techniques of intermittent control with adaptive aperiodicity to achieve the synchronization of a class of neural networks, modeled by stochastic systems. Some effective sufficient conditions are established for the realization of synchronization of the underlying network. Numerical simulations of two examples are provided to illustrate the theoretical results obtained in the paper.

Absorption in mind-wandering (MW) may worsen our mood and can cause psychological disorders. Researchers indicate the possibility that meta-awareness of MW prevents these mal-effects and enhances favorable consequences of MW, such as boosting creativity; thus, meta-awareness has attracted psychological and clinical attention. However, few studies have investigated the nature of meta-awareness of MW, because there has been no method to isolate and operate this ability. Therefore, we propose a new approach to manipulate the ability of meta-awareness. We used Pavlovian conditioning, tying to it an occurrence of MW and a neutral tone sound inducing the meta-awareness of MW. To perform paired presentations of the unconditioned stimulus (neutral tone) and the conditioned stimulus (perception accompanying MW), we detected participants' natural occurrence of MW via electroencephalogram and a machine-learning estimation method. The double-blinded randomized controlled trial with 37 participants found that a single 20-min conditioning session significantly increased the meta-awareness of MW as assessed by behavioral and neuroscientific measures. The core protocol of the proposed method is real-time feedback on participants' neural information, and in that sense, we can refer to it as neurofeedback. However, there are some differences from typical neurofeedback protocols, and we discuss them in this paper. Our novel classical conditioning is expected to contribute to future research on the modulation effect of meta-awareness on MW.

Deep Neural Networks are widely used for solving complex problems in several scientific areas, such as speech recognition, machine translation, image analysis. The strategies employed to investigate their theoretical properties mainly rely on Euclidean geometry, but in the last years new approaches based on Riemannian geometry have been developed. Motivated by some open problems, we study a particular sequence of maps between manifolds, with the last manifold of the sequence equipped with a Riemannian metric. We investigate the structures induced through pullbacks on the other manifolds of the sequence and on some related quotients. In particular, we show that the pullbacks of the final Riemannian metric to any manifolds of the sequence is a degenerate Riemannian metric inducing a structure of pseudometric space. We prove that the Kolmogorov quotient of this pseudometric space yields a smooth manifold, which is the base space of a particular vertical bundle. We investigate the theoretical properties of the maps of such sequence, eventually we focus on the case of maps between manifolds implementing neural networks of practical interest and we present some applications of the geometric framework we introduced in the first part of the paper.

To develop an efficient brain-computer interface (BCI) system, electroencephalography (EEG) measures neuronal activities in different brain regions through electrodes. Many EEG-based motor imagery (MI) studies do not make full use of brain network topology. In this paper, a deep learning framework based on a modified graph convolution neural network (M-GCN) is proposed, in which temporal-frequency processing is performed on the data through modified S-transform (MST) to improve the decoding performance of original EEG signals in different types of MI recognition. MST can be matched with the spatial position relationship of the electrodes. This method fusions multiple features in the temporal-frequency-spatial domain to further improve the recognition performance. By detecting the brain function characteristics of each specific rhythm, EEG generated by imaginary movement can be effectively analyzed to obtain the subjects' intention. Finally, the EEG signals of patients with spinal cord injury (SCI) are used to establish a correlation matrix containing EEG channel information, the M-GCN is employed to decode relation features. The proposed M-GCN framework has better performance than other existing methods. The accuracy of classifying and identifying MI tasks through the M-GCN method can reach 87.456%. After 10-fold cross-validation, the average accuracy rate is 87.442%, which verifies the reliability and stability of the proposed algorithm. Furthermore, the method provides effective rehabilitation training for patients with SCI to partially restore motor function.

Group convolution has been widely used in order to reduce the computation time of convolution, which takes most of the training time of convolutional neural networks. However, it is well known that a large number of groups significantly reduce the performance of group convolution. In this paper, we propose a new convolution methodology called ``two-level'' group convolution that is robust with respect to the increase of the number of groups and suitable for multi-GPU parallel computation. We first observe that the group convolution can be interpreted as a one-level block Jacobi approximation of the standard convolution, which is a popular notion in the field of numerical analysis. In numerical analysis, there have been numerous studies on the two-level method that introduces an intergroup structure that resolves the performance degradation issue without disturbing parallel computation. Motivated by these, we introduce a coarse-level structure which promotes intergroup communication without being a bottleneck in the group convolution. We show that all the additional work induced by the coarse-level structure can be efficiently processed in a distributed memory system. Numerical results that verify the robustness of the proposed method with respect to the number of groups are presented. Moreover, we compare the proposed method to various approaches for group convolution in order to highlight the superiority of the proposed method in terms of execution time, memory efficiency, and performance.

Improving existing neural network architectures can involve several design choices such as manipulating the loss functions, employing a diverse learning strategy, exploiting gradient evolution at training time, optimizing the network hyper-parameters, or increasing the architecture depth. The latter approach is a straightforward solution, since it directly enhances the representation capabilities of a network; however, the increased depth generally incurs in the well-known vanishing gradient problem. In this paper, borrowing from different methods addressing this issue, we introduce an interlaced multi-task learning strategy, defined SIRe, to reduce the vanishing gradient in relation to the object classification task. The presented methodology directly improves a convolutional neural network (CNN) by preserving information from the input image through interlaced auto-encoders (AEs), and further refines the base network architecture by means of skip and residual connections. To validate the presented methodology, a simple CNN and various implementations of famous networks are extended via the SIRe strategy and extensively tested on five collections, i.e., MNIST, Fashion-MNIST, CIFAR-10, CIFAR-100, and Caltech-256; where the SIRe-extended architectures achieve significantly increased performances across all models and datasets, thus confirming the presented approach effectiveness.

Random Forest is an ensemble of decision trees based on the bagging and random subspace concepts. As suggested by Breiman, the strength of unstable learners and the diversity among them are the ensemble models' core strength. In this paper, we propose two approaches known as oblique and rotation double random forests. In the first approach, we propose rotation based double random forest. In rotation based double random forests, transformation or rotation of the feature space is generated at each node. At each node different random feature subspace is chosen for evaluation, hence the transformation at each node is different. Different transformations result in better diversity among the base learners and hence, better generalization performance. With the double random forest as base learner, the data at each node is transformed via two different transformations namely, principal component analysis and linear discriminant analysis. In the second approach, we propose oblique double random forest. Decision trees in random forest and double random forest are univariate, and this results in the generation of axis parallel split which fails to capture the geometric structure of the data. Also, the standard random forest may not grow sufficiently large decision trees resulting in suboptimal performance. To capture the geometric properties and to grow the decision trees of sufficient depth, we propose oblique double random forest. The oblique double random forest models are multivariate decision trees. At each non-leaf node, multisurface proximal support vector machine generates the optimal plane for better generalization performance. Also, different regularization techniques (Tikhonov regularization, axis-parallel split regularization, Null space regularization) are employed for tackling the small sample size problems in the decision trees of oblique double random forest. The proposed ensembles of decision trees produce trees with bigger size compared to the standard ensembles of decision trees as bagging is used at each non-leaf node which results in improved performance. The evaluation of the baseline models and the proposed oblique and rotation double random forest models is performed on benchmark 121 UCI datasets and real-world fisheries datasets. Both statistical analysis and the experimental results demonstrate the efficacy of the proposed oblique and rotation double random forest models compared to the baseline models on the benchmark datasets.

In this paper we present ExSpliNet, an interpretable and expressive neural network model. The model combines ideas of Kolmogorov neural networks, ensembles of probabilistic trees, and multivariate B-spline representations. We give a probabilistic interpretation of the model and show its universal approximation properties. We also discuss how it can be efficiently encoded by exploiting B-spline properties. Finally, we test the effectiveness of the proposed model on synthetic approximation problems and classical machine learning benchmark datasets.

Square matrices appear in many machine learning problems and models. Optimization over a large square matrix is expensive in memory and in time. Therefore an economic approximation is needed. Conventional approximation approaches factorize the square matrix into a number matrices of much lower ranks. However, the low-rank constraint is a performance bottleneck if the approximated matrix is intrinsically high-rank or close to full rank. In this paper, we propose to approximate a large square matrix with a product of sparse full-rank matrices. In the approximation, our method needs only N(logN)

Despite considerable progress in the field of automatic multi-target tracking, several problems such as data association remained challenging. On the other hand, cognitive studies have reported that humans can robustly track several objects simultaneously. Such circumstances happen regularly in daily life, and humans have evolved to handle the associated problems. Accordingly, using brain-inspired processing principles may contribute to significantly increase the performance of automatic systems able to follow the trajectories of multiple objects. In this paper, we propose a multiple-object tracking algorithm based on dynamic neural field theory which has been proven to provide neuro-plausible processing mechanisms for cognitive functions of the brain. We define several input neural fields responsible for representing previous location and orientation information as well as instantaneous linear and angular speed of the objects in successive video frames. Image processing techniques are applied to extract the critical object features including target location and orientation. Two prediction fields anticipate the objects' locations and orientations in the upcoming frame after receiving excitatory and inhibitory inputs from the input fields in a feed-forward architecture. This information is used in the data association and labeling process. We tested the proposed algorithm on a zebrafish larvae segmentation and tracking dataset and an ant-tracking dataset containing non-rigid objects with spiky movements and frequently occurring occlusions. The results showed a significant improvement in tracking metrics compared to state-of-the-art algorithms., This work has been supported by the Center for International Scientific Studies and Collaboration (CISSC), Ministry of Science Research and Technology, Iran, Grant No. 1483.

In a competitive game scenario, a set of agents have to learn decisions that maximize their goals and minimize their adversaries' goals at the same time. Besides dealing with the increased dynamics of the scenarios due to the opponents' actions, they usually have to understand how to overcome the opponent's strategies. Most of the common solutions, usually based on continual learning or centralized multi-agent experiences, however, do not allow the development of personalized strategies to face individual opponents. In this paper, we propose a novel model composed of three neural layers that learn a representation of a competitive game, learn how to map the strategy of specific opponents, and how to disrupt them. The entire model is trained online, using a composed loss based on a contrastive optimization, to learn competitive and multiplayer games. We evaluate our model on a pokemon duel scenario and the four-player competitive Chef's Hat card game. Our experiments demonstrate that our model achieves better performance when playing against offline, online, and competitive-specific models, in particular when playing against the same opponent multiple times. We also present a discussion on the impact of our model, in particular on how well it deals with on specific strategy learning for each of the two scenarios.

This paper proposes a new hierarchical approach to learning rate adaptation in gradient methods, called learning rate optimization (LRO). LRO formulates the learning rate adaption problem as a hierarchical optimization problem that minimizes the loss function with respect to the learning rate for current model parameters and gradients. Then, LRO optimizes the learning rate based on the alternating direction method of multipliers (ADMM). In the process of this learning rate optimization, LRO does not require any second-order information and probabilistic model, so it is highly efficient. Furthermore, LRO does not require any additional hyperparameters when compared to the vanilla gradient method with the simple exponential learning rate decay. In the experiments, we integrated LRO with vanilla SGD and Adam. Then, we compared their optimization performance with the state-of-the-art learning rate adaptation methods and also the most commonly-used adaptive gradient methods. The SGD and Adam with LRO outperformed all the competitors on the benchmark datasets in image classification tasks.

Optimal stochastic controller pushes the closed-loop behavior as close as possible to the desired one. The fully probabilistic design (FPD) uses probabilistic description of the desired closed loop and minimizes Kullback-Leibler divergence of the closed-loop description to the desired one. Practical exploitation of the fully probabilistic design control theory continues to be hindered by the computational complexities involved in numerically solving the associated stochastic dynamic programming problem. In particular very hard multivariate integration and an approximate interpolation of the involved multivariate functions. This paper proposes a new fully probabilistic contro algorithm that uses the adaptive critic methods to circumvent the need for explicitly evaluating the optimal value function, thereby dramatically reducing computational requirements. This is a main contribution of this short paper.

Background noise and scale variation are common problems that have been long recognized in crowd counting. Humans glance at a crowd image and instantly know the approximate number of human and where they are through attention the crowd regions and the congestion degree of crowd regions with a global receptive field. Hence, in this paper, we propose a novel feedback network with Region-Aware block called RANet by modeling human's Top-Down visual perception mechanism. Firstly, we introduce a feedback architecture to generate priority maps that provide prior about candidate crowd regions in input images. The prior enables the RANet pay more attention to crowd regions. Then we design Region-Aware block that could adaptively encode the contextual information into input images through global receptive field. More specifically, we scan the whole input images and its priority maps in the form of column vector to obtain a relevance matrix estimating their similarity. The relevance matrix obtained would be utilized to build global relationships between pixels. Our method outperforms state-of-the-art crowd counting methods on several public datasets.

In this paper, a novel data-driven method for weight initialization of Multilayer Perceptrons and Convolutional Neural Networks based on discriminant learning is proposed. The approach relaxes some of the limitations of competing data-driven methods, including unimodality assumptions, limitations on the architectures related to limited maximal dimensionalities of the corresponding projection spaces, as well as limitations related to high computational requirements due to the need of eigendecomposition on high-dimensional data. We also consider assumptions of the method on the data and propose a way to account for them in a form of a new normalization layer. The experiments on three large-scale image datasets show improved accuracy of the trained models compared to competing random-based and data-driven weight initialization methods, as well as better convergence properties in certain cases. publishedVersion

Modifying the existing models of classifiers’ operation is primarily aimed at increasing the effectiveness as well as minimizing the training time. An additional advantage is the ability to quickly implement a given solution to the real needs of the market. In this paper, we propose a method that can implement various classifiers using the federated learning concept and taking into account parallelism. Also, an important element is the analysis and selection of the best classifier depending on its reliability found for separated datasets extended by new, augmented samples. The proposed augmentation technique involves image processing techniques, neural architectures, and heuristic methods and improves the operation in federated learning by increasing the role of the server. The proposition has been presented and tested for the fruit image classification problem. The conducted experiments have shown that the described technique can be very useful as an implementation method even in the case of a small database. Obtained results are discussed concerning the advantages and disadvantages in the context of practical application like higher accuracy.

Counting objects in images is a very time-consuming task for humans that yields to errors caused by repetitiveness and boredom. In this paper, we present a novel object counting method that, unlike most of the recent works that focus on the regression of a density map, performs the counting procedure by localizing each single object. This key difference allows us to provide not only an accurate count but the position of every counted object, information that can be critical in some areas such as precision agriculture. The method is designed in two steps: first, a CNN is in charge of mapping arbitrary objects to blob-like structures. Then, using a Laplacian of Gaussian (LoG) filter, we are able to gather the position of all detected objects. We also propose a semi-adversarial training procedure that, combined with the former design, improves the result by a large margin. After evaluating the method on two public benchmarks of isometric objects, we stay on par with the state of the art while being able to provide extra position information.

Emotion classification based on neurophysiology signals has being a challenging issue in the literature. Recent neuroscience findings suggest that brain network structure underlying the different emotions provide a window in understanding human affection. In this paper, we propose a novel method to capture the distinct minimum spanning tree (MST) topology underpinning the different emotions. Specifically, we propose a hierarchical aggregation-based graph neural network to investigate the MST structure in emotion recognition. Extensive experiments on the public available DEAP dataset demonstrate the superior performance of the model in emotion classification as compared to existing methods. In addition, the results show that the theta, lower beta and gamma frequency band network information are more sensitive to emotions, suggesting a multi-frequency interaction in emotion processing.

In this paper, the issue of synchronization is investigated for coupled neural networks subject to stochastic deception attacks. Firstly, a general differential inequality with delayed impulses is given. Then, the established differential inequality is further extended to the case of delayed stochastic impulses, in which both the impulsive instants and impulsive intensity are stochastic. Secondly, by modeling the stochastic discrete-time deception attacks as stochastic impulses, synchronization criteria of the coupled neural networks under the corresponding attacks are given. Finally, two numerical examples are provided to demonstrate the correctness of the theoretical results.

The recurrent network architecture is a widely used model in sequence modeling, but its serial dependency hinders the computation parallelization, which makes the operation inefficient. The same problem was encountered in serial adder at the early stage of digital electronics. In this paper, we discuss the similarities between recurrent neural network (RNN) and serial adder. Inspired by carry-lookahead adder, we introduce carry-lookahead module to RNN, which makes it possible for RNN to run in parallel. Then, we design the method of parallel RNN computation, and finally Carry-lookahead RNN (CL-RNN) is proposed. CL-RNN takes advantages in parallelism and flexible receptive field. Through a comprehensive set of tests, we verify that CL-RNN can perform better than existing typical RNNs in sequence modeling tasks which are specially designed for RNNs. Code and models are available at: https://github.com/WinnieJiangHW/Carry-lookahead_RNN.

Recurrent Neural Networks with Long Short-Term Memory (LSTM) make use of gating mechanisms to mitigate exploding and vanishing gradients when learning long-term dependencies. For this reason, LSTMs and other gated RNNs are widely adopted, being the standard de facto for many sequence modeling tasks. Although the memory cell inside the LSTM contains essential information, it is not allowed to influence the gating mechanism directly. In this work, we improve the gate potential by including information coming from the internal cell state. The proposed modification, named Working Memory Connection, consists in adding a learnable nonlinear projection of the cell content into the network gates. This modification can fit into the classical LSTM gates without any assumption on the underlying task, being particularly effective when dealing with longer sequences. Previous research effort in this direction, which goes back to the early 2000s, could not bring a consistent improvement over vanilla LSTM. As part of this paper, we identify a key issue tied to previous connections that heavily limits their effectiveness, hence preventing a successful integration of the knowledge coming from the internal cell state. We show through extensive experimental evaluation that Working Memory Connections constantly improve the performance of LSTMs on a variety of tasks. Numerical results suggest that the cell state contains useful information that is worth including in the gate structure., Accepted for publication in Neural Networks

In this paper, we propose a novel ensembling technique for deep neural networks, which is able to drastically reduce the required memory compared to alternative approaches. In particular, we propose to extract multiple sub-networks from a single, untrained neural network by solving an end-to-end optimization task combining differentiable scaling over the original architecture, with multiple regularization terms favouring the diversity of the ensemble. Since our proposal aims to detect and extract sub-structures, we call it Structured Ensemble. On a large experimental evaluation, we show that our method can achieve higher or comparable accuracy to competing methods while requiring significantly less storage. In addition, we evaluate our ensembles in terms of predictive calibration and uncertainty, showing they compare favourably with the state-of-the-art. Finally, we draw a link with the continual learning literature, and we propose a modification of our framework to handle continuous streams of tasks with a sub-linear memory cost. We compare with a number of alternative strategies to mitigate catastrophic forgetting, highlighting advantages in terms of average accuracy and memory., Comment: Article accepted at Neural Networks

Deep convolutional neural network compression has attracted lots of attention due to the need to deploy accurate models on resource-constrained edge devices. Existing techniques mostly focus on compressing networks for image-level classification, and it is not clear if they generalize well on network architectures for more challenging pixel-level tasks, e.g., dense crowd counting or semantic segmentation. In this paper, we propose an adaptive correlation-driven sparsity learning (ACSL) framework for channel pruning that outperforms state-of-the-art methods on both image-level and pixel-level tasks. In our ACSL framework, we first quantify the data-dependent channel correlation information with a channel affinity matrix. Next, we leverage these inter-dependencies to induce sparsity into the channels with the introduced adaptive penalty strength. After removing the redundant channels, we obtain compact and efficient models, which have significantly less number of parameters while maintaining comparable performance with the original models. We demonstrate the advantages of our proposed approach on three popular vision tasks, i.e., dense crowd counting, semantic segmentation, and image-level classification. The experimental results demonstrate the superiority of our framework. In particular, for crowd counting on the Mall dataset, the proposed ACSL framework is able to reduce up to 94% parameters (VGG16-Decoder) and 84% FLOPs (ResNet101), while maintaining the same performance of (at times outperforming) the original model. National Research Foundation (NRF) This research project is supported by the National Research Foundation Singapore under its Campus for Research Excellence and Technological Enterprise (CREATE) programme with the Technical University of Munich at TUMCREATE. Prof. Siew-Kei Lam is partially supported under the RIE2020 Industry Alignment Fund – Industry Collaboration Projects (IAFICP) Funding Initiative, as well as cash and in-kind contribution from Singapore Telecommunications Limited (Singtel), through Singtel Cognitive and Artificial Intelligence Lab for Enterprises (SCALE@NTU).

In this paper, we propose a novel transductive pseudo-labeling based method for deep semi-supervised image recognition. Inspired from the superiority of pseudo labels inferred by label propagation compared with those inferred from network, we argue that information flow from labeled data to unlabeled data should be kept noiseless and with minimum loss. Previous research works use scarce labeled data for feature learning and solely consider the relationship between two feature vectors to construct the similarity graph in feature space, which causes two problems that ultimately lead to noisy and incomplete information flow from labeled data to unlabeled data. The first problem is that the learned feature mapping is highly likely to be biased and can easily over-fit noise. The second problem is the loss of local geometry information in feature space during label propagation. Accordingly, we firstly propose to incorporate self-supervised learning into feature learning for cleaner information flow in feature space during subsequent label propagation. Secondly, we propose to use reconstruction concept to measure pairwise similarity in feature space, such that local geometry information can be preserved. Ablation study confirms synergistic effects from features learned with self-supervision and similarity graph with local geometry preserving. Extensive experiments conducted on benchmark datasets have verified the effectiveness of our proposed method.

We use a recently developed synchronous Spiking Neural Network (SNN) model to study the problem of learning hierarchically-structured concepts. We introduce an abstract data model that describes simple hierarchical concepts. We define a feed-forward layered SNN model, with learning modeled using Oja’s local learning rule, a well known biologically-plausible rule for adjusting synapse weights. We define what it means for such a network to recognize hierarchical concepts; our notion of recognition is robust, in that it tolerates a bounded amount of noise. Then, we present a learning algorithm by which a layered network may learn to recognize hierarchical concepts according to our robust definition. We analyze correctness and performance rigorously; the amount of time required to learn each concept, after learning all of the sub-concepts, is approximately O 1 η k lmax log ( k ) + 1 ɛ + b log ( k ) , where k is the number of sub-concepts per concept, lmax is the maximum hierarchical depth, η is the learning rate, ɛ describes the amount of uncertainty allowed in robust recognition, and b describes the amount of weight decrease for “irrelevant” edges. An interesting feature of this algorithm is that it allows the network to learn sub-concepts in a highly interleaved manner. This algorithm assumes that the concepts are presented in a noise-free way; we also extend these results to accommodate noise in the learning process. Finally, we give a simple lower bound saying that, in order to recognize concepts with hierarchical depth two with noise-tolerance, a neural network should have at least two layers. The results in this paper represent first steps in the theoretical study of hierarchical concepts using SNNs. The cases studied here are basic, but they suggest many directions for extensions to more elaborate and realistic cases.

Routing methods in capsule networks often learn a hierarchical relationship for capsules in successive layers, but the intra-relation between capsules in the same layer is less studied, while this intra-relation is a key factor for the semantic understanding in text data. Therefore, in this paper, we introduce a new capsule network with graph routing to learn both relationships, where capsules in each layer are treated as the nodes of a graph. We investigate strategies to yield adjacency and degree matrix with three different distances from a layer of capsules, and propose the graph routing mechanism between those capsules. We validate our approach on five text classification datasets, and our findings suggest that the approach combining bottom-up routing and top-down attention performs the best. Such an approach demonstrates generalization capability across datasets. Compared to the state-of-the-art routing methods, the improvements in accuracy in the five datasets we used were 0.82, 0.39, 0.07, 1.01, and 0.02, respectively.

Inspired by the success of classical neural networks, there has been tremendous effort to develop classical effective neural networks into quantum concept. In this paper, a novel hybrid quantum–classical neural network with deep residual learning (Res-HQCNN) is proposed. We firstly analyse how to connect residual block structure with a quantum neural network, and give the corresponding training algorithm. At the same time, the advantages and disadvantages of transforming deep residual learning into quantum concept are provided. As a result, the model can be trained in an end-to-end fashion, analogue to the backpropagation in classical neural networks. To explore the effectiveness of Res-HQCNN , we perform extensive experiments for quantum data with or without noisy on classical computer. The experimental results show the Res-HQCNN performs better to learn an unknown unitary transformation and has stronger robustness for noisy data, when compared to state of the arts. Moreover, the possible methods of combining residual learning with quantum neural networks are also discussed.

Graph autoencoders (AE) and variational autoencoders (VAE) are powerful node embedding methods, but suffer from scalability issues. In this paper, we introduce FastGAE, a general framework to scale graph AE and VAE to large graphs with millions of nodes and edges. Our strategy, based on an effective stochastic subgraph decoding scheme, significantly speeds up the training of graph AE and VAE while preserving or even improving performances. We demonstrate the effectiveness of FastGAE on various real-world graphs, outperforming the few existing approaches to scale graph AE and VAE by a wide margin., Accepted for publication in Elsevier's Neural Networks journal

An approach to the time-accurate prediction of chaotic solutions is by learning temporal patterns from data. Echo State Networks (ESNs), which are a class of Reservoir Computing, can accurately predict the chaotic dynamics well beyond the predictability time. Existing studies, however, also showed that small changes in the hyperparameters may markedly affect the network's performance. The aim of this paper is to assess and improve the robustness of Echo State Networks for the time-accurate prediction of chaotic solutions. The goal is three-fold. First, we investigate the robustness of routinely used validation strategies. Second, we propose the Recycle Validation, and the chaotic versions of existing validation strategies, to specifically tackle the forecasting of chaotic systems. Third, we compare Bayesian optimization with the traditional Grid Search for optimal hyperparameter selection. Numerical tests are performed on two prototypical nonlinear systems that have both chaotic and quasiperiodic solutions. Both model-free and model-informed Echo State Networks are analysed. By comparing the network's robustness in learning chaotic versus quasiperiodic solutions, we highlight fundamental challenges in learning chaotic solutions. The proposed validation strategies, which are based on the dynamical systems properties of chaotic time series, are shown to outperform the state-of-the-art validation strategies. Because the strategies are principled-they are based on chaos theory such as the Lyapunov time-they can be applied to other Recurrent Neural Networks architectures with little modification. This work opens up new possibilities for the robust design and application of Echo State Networks, and Recurrent Neural Networks, to the time-accurate prediction of chaotic systems., Comment: 29 pages, 25 figures

Event cameras contain emerging, neuromorphic vision sensors that capture local-light​ intensity changes at each pixel, generating a stream of asynchronous events. This way of acquiring visual information constitutes a departure from traditional frame-based cameras and offers several significant advantages — low energy consumption, high temporal resolution, high dynamic range and low latency. Driver monitoring systems (DMS) are in-cabin safety systems designed to sense and understand a drivers physical and cognitive state. Event cameras are particularly suited to DMS due to their inherent advantages. This paper proposes a novel method to simultaneously detect and track faces and eyes for driver monitoring. A unique, fully convolutional recurrent neural network architecture is presented. To train this network, a synthetic event-based dataset is simulated with accurate bounding box annotations, called Neuromorphic-HELEN. Additionally, a method to detect and analyse drivers’ eye blinks is proposed, exploiting the high temporal resolution of event cameras. Behaviour of blinking provides greater insights into a driver level of fatigue or drowsiness. We show that blinks have a unique temporal signature that can be better captured by event cameras.

Traditional active noise control (ANC) methods are based on adaptive signal processing with the least mean square algorithm as the foundation. They are linear systems and do not perform satisfactorily in the presence of nonlinear distortions. In this paper, we formulate ANC as a supervised learning problem and propose a deep learning approach, called deep ANC, to address the nonlinear ANC problem. The main idea is to employ deep learning to encode the optimal control parameters corresponding to different noises and environments. A convolutional recurrent network (CRN) is trained to estimate the real and imaginary spectrograms of the canceling signal from the reference signal so that the corresponding anti-noise can eliminate or attenuate the primary noise in the ANC system. Large-scale multi-condition training is employed to achieve good generalization and robustness against a variety of noises. The deep ANC method can be trained to achieve active noise cancellation no matter whether the reference signal is noise or noisy speech. In addition, a delay-compensated strategy is introduced to solve the potential latency problem of ANC systems. Experimental results show that deep ANC is effective for wideband noise reduction and generalizes well to untrained noises. Moreover, the proposed method can achieve ANC within a quiet zone and is robust against variations in reference signals.

Ensemble learning methods combine multiple models to improve performance by exploiting their diversity. The success of these approaches relies heavily on the dissimilarity of the base models forming the ensemble. This diversity can be achieved in many ways, with well-known examples including bagging and boosting. It is the diversity of the models within an ensemble that allows the ensemble to correct the errors made by its members, and consequently leads to higher classification or regression performance. A mistake made by a base model can only be rectified if other members behave differently on that particular instance, and provide the aggregator with enough information to make an informed decision. On the contrary, lack of diversity not only lowers model performance, but also wastes computational resources. Nevertheless, in the current state of the art ensemble approaches, there is no guarantee on the level of diversity achieved, and no mechanism ensuring that each member will learn a different decision boundary from the others. In this paper, we propose a parallel orthogonal deep learning architecture in which diversity is enforced by design, through imposing an orthogonality constraint. Multiple deep neural networks are created, parallel to each other. At each parallel layer, the outputs of different base models are subject to Gram–Schmidt orthogonalization. We demonstrate that this approach leads to a high level of diversity from two perspectives. First, the models make different errors on different parts of feature space, and second, they exhibit different levels of uncertainty in their decisions. Experimental results confirm the benefits of the proposed method, compared to standard deep learning models and well-known ensemble methods, in terms of diversity and, as a result, classification performance.

El pdf del artículo es la versión post-print., This paper presents a modification to the original ART 1 algorithm (Carpenter and Grossberg, 1987a, A massively parallel architecture for a self-organizing neural pattern recognition machine, Computer Vision, Graphics, and Image Processing, 37, 54–115) that is conceptually similar, can be implemented in hardware with less sophisticated building blocks, and maintains the computational capabilities of the originally proposed algorithm. This modified ART 1 algorithm (which we will call here ART 1m) is the result of hardware motivated simplifications investigated during the design of an actual ART 1 chip [Serrano-Gotarredona et al., 1994, Proc. 1994 IEEE Int. Conf. Neural Networks (Vol. 3, pp. 1912–1916); Serrano-Gotarredona and Linares-Barranco, 1996, IEEE Trans. VLSI Systems, (in press)]. The purpose of this paper is simply to justify theoretically that the modified algorithm preserves the computational properties of the original one and to study the difference in behavior between the two approaches.