Advances in Brain Inspired Cognitive Systems

Encouraged by the success of conventional GradientNet and recently-proposed ZhangNet for online equality-constrained quadratic programming problem, an improved recurrent network and its electronic implementation are firstly proposed and developed in this paper. Exploited in the primal form of quadratic programming with linear equality constraints, the proposed neural model can solve the problem effectively. Moreover, compared to the existing recurrent networks, i.e., GradientNet (GN) and ZhangNet (ZN), our model can theoretically guarantee superior global exponential convergence performance. Robustness performance of our such neural model is also analysed under a large model implementation error, with the upper bound of stead-state solution error estimated. Simulation results demonstrate theoretical analysis on the proposed model for online equality-constrained quadratic programming.

Mirror Self-Recognition is a well accepted test to identify whether an animal is with self-consciousness. Mirror neuron system is believed to be one of the most important biological foundation for Mirror Self-Recognition. Inspired by the biological mirror neuron system of the mammalian brain, we propose a Brain-inspired Robot Mirror Neuron System Model (Robot-MNS-Model) and we apply it to humanoid robots for mirror self-recognition. This model evaluates the similarity between the actual movements of robots and their visual perceptions. The association for self-recognition is supported by STDP learning which connects the correlated visual perception and motor control. The model is evaluated on self-recognition mirror test for 3 humanoid robots. Each robot has to decide which one is itself after a series of random movements facing a mirror. The results show that with the proposed model, multiple robots can pass the self-recognition mirror test at the same time, which is a step forward towards robot self-consciousness.

Deep Learning (DL) is capable of excavating features hidden deep in complex data. In this paper, we introduce hierarchical convolutional neural networks (HCNN) to implement the EEG-based emotion classifier (positive, negative and neutral) in a movie-watching task. Differential Entropy (DE) is calculated as features at certain time interval for each channel. We organize features from different channels into two dimensional maps to train HCNN classifier. This approach extracts features contained in the spatial topology of electrodes directly, which is often neglected by the widely-used one-dimensional models. The performance of HCNN was compared with one-dimensional deep model SAE (Stacked Autoencoder), as well as traditional shallow models SVM and KNN. We find that HCNN (88.2% ± 3.5%) is better than SAE (85.4% ± 8.1%), and deep models are more favorable in emotion recognition BCI (Brain-computer Interface) system than shallow models. Moreover, we show that models learned on one person is hard to transfer to others and the individual difference in EEG emotion-related signal is significant among peoples. Finally, we find Beta and Gamma (rather than Delta, Theta and Alpha) waves play the key role in emotion recognition.

Recent progress in neural learning demonstrated that machines can do well in regularized tasks, e.g., the game of Go. However, artistic activities such as poem generation are still widely regarded as human’s special capability. In this paper, we demonstrate that a simple neural model can imitate human in some tasks of art generation. We particularly focus on traditional Chinese poetry, and show that machines can do as well as many contemporary poets and weakly pass the Feigenbaum Test, a variant of Turing test in professional domains.Our method is based on an attention-based recurrent neural network, which accepts a set of keywords as the theme and generates poems by looking at each keyword during the generation. A number of techniques are proposed to improve the model, including character vector initialization, attention to input and hybrid-style training. Compared to existing poetry generation methods, our model can generate much more theme-consistent and semantic-rich poems.

A universal unanswered question in neuroscience and machine learning is whether computers can decode the patterns of the human brain. Multi-Voxels Pattern Analysis (MVPA) is a critical tool for addressing this question. However, there are two challenges in the previous MVPA methods, which include decreasing sparsity and noises in the extracted features and increasing the performance of prediction. In overcoming mentioned challenges, this paper proposes Anatomical Pattern Analysis (APA) for decoding visual stimuli in the human brain. This framework develops a novel anatomical feature extraction method and a new imbalance AdaBoost algorithm for binary classification. Further, it utilizes an Error-Correcting Output Codes (ECOC) method for multi-class prediction. APA can automatically detect active regions for each category of the visual stimuli. Moreover, it enables us to combine homogeneous datasets for applying advanced classification. Experimental studies on 4 visual categories (words, consonants, objects and scrambled photos) demonstrate that the proposed approach achieves superior performance to state-of-the-art methods.

This paper reviews the state of the art techniques for designing next generation CDSSs. CDSS can aid physicians and radiologists to better analyse and treat patients by combining their respective clinical expertise with complementary capabilities of the computers. CDSSs comprise many techniques from inter-desciplinary fields of medical image acquisition, image processing and pattern recognition, neural perception and pattern classifiers for medical data organization, and finally, analysis and optimization to enhance overall system performance. This paper discusses some of the current challenges in designing an efficient CDSS as well as some of the latest techniques that have been proposed to meet these challenges, primarily, by finding informative patterns in the medical dataset, analysing them and building a descriptive model of the object of interest, thus aiding in enhanced medical diagnosis.

We propose a new tone mapping model to render high dynamic range (HDR) images in limited dynamic range devices in this paper. This neural network model is inspired by the retinal information processing mechanisms of the biological visual system, including the adaptive gap junction between horizontal cells (HCs), the negative HC-cone feedback pathway, and the center-surround antagonistic receptive fields of bipolar cells (BCs). The key novelty of the proposed model lies in the adaptive adjustment of the receptive field size of HCs based on the local brightness, which simulates the dynamic gap junction between HCs. This enables the brightness of distinct regions to be recovered into clearly visible ranges while reducing halo artifacts common to other methods. The BCs serve to enhance the local contrast with their center-surround RF structure. By comparing with the state-of-the-art tone mapping methods qualitatively and quantitatively, our method shows competitive performance in term of improving details in both dark and bright areas.

In this paper, we propose a new approach for unsupervised learning using autoencoders with drop strategy (DrAE). Different from Explicit Regularized Autoencoders (ERAE), DrAE has no any additionally explicit regularization term to the cost function. A serial of drop strategies are exploited in the training phase of autoencoders for robust feature representation, such as dropout, dropConnect, denoising, winner-take-all, local winner-take-all. When training DrAE, subset of units or weights are set to zero. The results of our experiments on the MNIST dataset show that the performance of DrAE is better or comparative to ERAE.

Detection of unanticipated and rare events refers to a process of identifying an occasional target (oddball) stimulus from a regular trail of standard stimuli based on brain wave signals. It is the premise of human event-related potential (ERP) applications, a significant research topic in brain computer interfaces. The focus of this paper is to investigate whether unanticipated and rare visual and auditory events are detectable from EEG signals. In order to achieve this, an exploratory experiment is conducted. A novel pairwise comparison neural network approach to detect those unanticipated and rare visual and auditory events from EEG signals is introduced. Results indicate that the change in EEG signals caused by unanticipated rare events is detectable; a piece of finding that opens opportunities for ERP-based applications.

State-of-the-art deep neural networks (DNNs) have greatly improved the performance of facial landmarks detection. However, DNN models usually have a large number of parameters, which leads to high computational complexity and memory cost. To address this problem, we propose a method to compress large deep neural networks, which includes three steps. (1) Importance-based neuron pruning: compared with traditional connection pruning, we introduce weights correlations to prune unimportant neurons, which can reduce index storage and inference computation costs. (2) Product quantization: further use of product quantization helps to enforce weights sharing, which stores fewer cluster indexes and codebooks than scalar quantization. (3) Network retraining: to reduce training difficulty and performance degradation, we iteratively retrain the network, compressing one layer at a time. Experiments of compressing a VGG-like model for facial landmarks detection demonstrate that the proposed method achieves 26x compression of the model with 1.5% performance degradation.

Visual saliency detection is useful for applications as object recognition, resizing and image segmentation. It is a challenge to detect the most important scene from the input image. In this paper, we present a new method to get saliency map. First, we evaluate the salience value of each region by global contrast based spatial and color feature. Second, the salience values of the first stage are used to optimize the background and foreground queries (seeds), and then manifold ranking is employed to compute two phase saliency maps. Finally, the final saliency map is got by combining the two saliency map. Experiment results on four datasets indicate the significantly improved accuracy of the proposed algorithm in comparison with eight state-of-the-art approaches.

Inspired by biological spike information processing and the multiple brain region coordination mechanism, we propose an autonomous spiking neural network model for decision making. The proposed model is an expansion of the basal ganglia circuitry with automatic environment perception. It automatically constructs environmental states from image inputs. Contributions of this investigation can be summarized as the following: (1) In our model, the simplified Hodgkin-Huxley computing model is developed to achieve calculation efficiency closed to the LIF model and is used to obtain and test the ionic level properties in cognition. (2) A spike based motion perception mechanism is proposed to extract key elements for learning process from raw pixels without large amount of training. We apply our model in the “flappy bird” game and after dozens of training times, it can automatically generate rules to play well in the game. Besides, our model simulates cognitive defects when blocking some of sodium or potassium ion channels in the Hodgkin-Huxley model and this can be considered as a computational exploration on the mechanisms of cognition deep into ionic level.

Classification of spatiotemporal events captured by neuromorphic vision sensors or event based cameras in which each pixel senses the luminance changes of related spatial location and produces a sequence of events, has been of great interest in recent years. In this paper, we find that the classification accuracy can be significantly improved by combing random forest (RF) classifier with pixel-wise features. RF is a statistical framework with high generalization accuracy and fast training time. We uncover that random forest could grow deep and tend to learn highly irregular patterns of spatiotemporal events with low bias, and thus it is more suitable for achieving the classification objective. The experimental results on MNIST-DVS dataset and AER Posture dataset show that the RF based classification approach in this work outperforms the state of art algorithms in both classification accuracy and computation time cost.

The selective visual attention mechanism in human visual system helps human to act efficiently when dealing with massive visual information. Over the last two decades, biologically inspired attention model has drawn lots of research attention and many models have been proposed. However, the top-down cues in human brain are still not fully understood, which makes top-down models not biologically plausible. This paper proposes an attention model containing both the bottom-up stage and top-down stage for the target detection from SAR (Synthetic Aperture Radar) images. The bottom-up stage is based on the biologically-inspired Itti model and is modified by taking fully into account the characteristic of SAR images. The top-down stage contains a novel learning strategy to make the full use of prior information. It is an extension of the bottom-up process and more biologically plausible. The experiments in this research aim to detect vehicles in different scenes to validate the proposed model by comparing with the well-known CFAR (constant false alarm rate) algorithm.

Clustering is one of the most challenging optimization problems. Many Swarm Intelligence techniques including Ant Colony optimization (ACO), Particle Swarm Optimization (PSO), and Honey Bee Optimization (HBO) have been used to solve clustering. Cat Swarm Optimization (CSO) is one of the newly proposed heuristics in swarm intelligence, which is generated by observing the behavior of cats, and has been used for clustering and numerical function optimization. CSO based clustering is dependent on a pre-specified value of K i.e. Number of Clusters. In this paper we have proposed a “Modified Cat Swam Optimization (MCSO)” heuristic to discover clusters based on the nature of data rather than user specified K. MCSO performs a data scan to determine the initial cluster centers. We have compared the results of MCSO with CSO to demonstrate the enhanced efficiency and accuracy of our proposed technique.

Large-scale deep neural models, e.g., deep neural networks (DNN) and recurrent neural networks (RNN), have demonstrated significant success in solving various challenging tasks of speech and language processing (SLP), including speech recognition, speech synthesis, document classification and question answering. This growing impact corroborates the neurobiological evidence concerning the presence of layer-wise deep processing in the human brain. On the other hand, sparse coding representation has also gained similar success in SLP, particularly in signal processing, demonstrating sparsity as another important neurobiological characteristic. Recently, research in these two directions is leading to increasing cross-fertlisation of ideas, thus a unified Sparse Deep or Deep Sparse learning framework warrants much attention. This paper aims to provide an overview of growing interest in this unified framework, and also outlines future research possibilities in this multi-disciplinary area.

Being one of the most direct expressions of human feelings, music becomes the best tool for investigating the relationship between emotion and cognition. This paper investigated the long-term spectrum evidence of electroencephalogram (EEG) activities elicited by music keys. The EEG signals were recorded in 21 healthy adults during the entire process of music listening. There were two major music episodes and two minor ones, each lasted two minutes. Considering the spectral characteristics: (1) During two minutes of music listening, the alpha band activities recovered rapidly, which were more obvious under major music; (2) while, in the high-frequency gamma band, the activities declined gradually which were more obvious under minor music. Taken together, these results give clear evidence for the time-course difference in the music key perception.

To adapt to the environment and survive, most animals can control their behaviors by making decisions. The process of decision making and responding according to changes in the environment is stable, sustainable, and learnable. Understanding how behaviors are regulated by neural circuits, and the encoding and decoding mechanisms from stimuli to responses, are important goals in neuroscience. A biologically plausible decision circuit consisting of computational neuron and synapse models and its learning mechanism are designed in this paper. The learning mechanism is based on two parts: first, effect of the punishment from the environment on the temporal correlations of neuron firings; second, spike timing dependent plasticity (STDP) of synapse. The decision circuit was used successfully to simulate the behavior of Drosophila exhibited in real experiments. In this paper, we place focus on the connections and interactions among excitatory and inhibitory neurons and try to give an explanation at a micro level (i.e. neurons and neural circuit) of how the observable decision making behavior is acquired and achieved.

Tinnitus is a kind of auditory disease characterized by an ongoing conscious perception of a sound in the absence of any external sound source. It is a common symptom for which no effective treatment exists. Though many non-invasive functional imaging modalities have been rapidly developed and applied to this field, yet, whether the EEG signal can be utilized to distinguish tinnitus patients from normal populations has not been investigated. In the present study, we perform a binary classification based on EEG signal to distinguish tinnitus patients from normal populations. In this study, 22 subjects are involved in the experiment with 15 of them being tinnitus patients and the others being normal controls. The collected EEG signals are preprocessed in frequency domain and well represented as features that depict each subject. Then the linear support vector machine is applied to classify the subjects. Satisfactory results have been achieved, where the accuracy of the classification could reach 90.91% in spite of the undeniable fact that the collected EEG signals contain noises. Accordingly, the present study reveals that the EEG signals can be utilized to distinguish tinnitus patients from normal populations, which could be regarded as an auxiliary therapy in tinnitus.

Passive brain-computer interface (BCI) can monitor cognitive function through physiological signals in human-machine system. This paper established a passive BCI based on functional near-infrared spectroscopy (fNIRS) to detect the sustained attentional load. Three levels of attentional load were adjusted by modifying the number of stimulate in feature-absence Continuous Performance Test (CPT) tasks. 15 healthy subjects were recruited in total, and 10 channels were measured in prefrontal cortex (PFC). Performance and NASA-TLX scales were also recorded as reference. The mean value of oxyhemoglobin and deoxyhemoglobin, signal slope, power spectrum and approximate entropy in 0–10 s were extracted from raw fNIRS signal for support vector machine (SVM) classification. The best performance features were selected by SVM-RFE algorithm. In conclusion over 80% average accuracy was achived between easy and hard attentional load, which demonstrated fNIRS can be a proposed method to detect sustained attention load for a passive BCI.

Incremental learning is a way relevant to human learning that utilizes samples in online sequence. In the paper, we propose an incremental learning method called Incremental Adaptive Learning Vector Quantization (IALVQ) which aims at classifying characters appearing in an online sequence with style consistency in local time periods. Such local consistency is present commonly in document images, in that the characters in a paragraph or text line are printed in the same font or written by the same person. Our IALVQ method updates the prototypes (parameters of classifier) incrementally to adapt to drifted concepts globally while utilize the style consistency locally. For style adaptation, a style transfer mapping (STM) matrix is calculated on a batch of samples of assumed same style. The STM matrix can be used both in training for prototypes updating and in testing for labels prediction. We consider supervised incremental learning and active incremental learning. In the latter way, class labels are attached only to samples that are assigned low confidence by the classifier. In our experiments on handwritten digits in the NIST Special Database 19, we evaluated the classification performance of IALVQ in two scenarios, interleaved test-then-train and style-specific classification. The results show that utilizing local style consistency can improve the accuracies of both two test scenarios, and for both supervised and active incremental learning modes.

Early detection and diagnosis of Hepatocellular Carcinoma (HCC) is the most discriminating step in liver cancer management. Image processing is primarily used, where fast and accurate Computed Tomography (CT) liver image segmentation is required for effective clinical studies and treatment plans. The purpose of this research is to develop an automated HCC detection and diagnosis system, able to work with HCC lesions from liver CT images, with maximum sensitivity and minimum specificity.Our proposed system carried out automated segmentation of HCC lesions from 3D liver CT images. First, based on chosen histogram thresholds, we create a mask to predict the segmentation area by exploiting prior knowledge of the location and shape. Next, we obtain a 3D HCC lesion using an appropriate combination of cancer area pixel density calculations, histogram analysis and morphological processing. To demonstrate the feasibility of our approach, we carried out a series of experiments using 31 CT cases, comprised of 18 HCC lesions and 13 non HCC lesions. The acquired CT images (in DICOM format) had 128 channels of 512 × 512 pixels, each with pixel space varying between 0.54 and 0.85. Simulation results showed 92.68% accuracy and a false positive incidence of 9.75%. These were also compared and validated against manual segmentation carried out by a radiologist and other widely used image segmentation methods.Fully automated HCC detection can be efficiently used to aid medical professionals in diagnosing HCC. A limitation of this research is that the performance was evaluated on a small dataset, which does not allow us to confirm robustness of this system. For future work, we will collect additional clinical and CT image data to ensure comprehensive evaluation and clinical validation. We also intend to apply this automated HCC detection and diagnosis system to Positron Emission Tomography (PET) and Magnetic Resonance Imaging (MRI) datasets, as well as adapting it for diagnosing different liver diseases using state-of-the-art feature extraction and selection, and machine learning classification techniques.

Millions of users world-wide are sharing content using the Peer-to-Peer (P2P) client network. While new innovations bring benefits, there are nevertheless some dangers associated with them. One of the main threats is P2P worms that can penetrate the network even from a single node and can then spread very quickly. Many attempts have been made in this domain to model the worm propagation behaviour, and yet no single model exists that can realistically model the process. Most researchers have considered disease epidemic models for modelling the worm propagation process. Such models are, however, based on strong assumptions which may not necessarily be valid in real-world scenarios. In this paper, a new biologically-inspired analytical model is proposed, one that considers configuration diversity, infection time lag, user-behaviour and node mobility as the important parameters that affect the worm propagation process. The model is flexible and can represent a network where all nodes are mobile or a heterogeneous network, where some nodes are static and others are mobile. A complete derivation of each of the factors is provided in the analytical model, and the results are benchmarked against recently reported analytical models. A comparative analysis of simulation results indeed shows that our proposed biologically-inspired model represents a more realistic picture of the worm propagation process, compared to the existing state-of-the-art analytical models.

Estimating the functional interactions and connections between brain regions to corresponding process in cognitive, behavioral and psychiatric domains are central pursuits for understanding the human connectome. Few studies have examined the effects of dynamic evolution on cognitive processing and brain activation using wavelet coherence in scalp electroencephalography (EEG) data. Aim of this study was to investigate the brain functional connectivity and dynamic programming model based on the wavelet coherence from EEG data and to evaluate a possible correlation between the brain connectivity architecture and cognitive evolution processing. Here, We present an accelerated dynamic programing algorithm that we found that spatially distributed regions coherence connection difference, for variation audio stimulation, dynamic programing model give the dynamic evolution processing in difference time and frequency. Such methodologies will be suitable for capturing the dynamic evolution of the time varying connectivity patterns that reflect certain cognitive tasks or brain pathologies.

This study proposes a new diagnostic approach based on application of machine learning techniques to anthropometric patient features in order to create a predictive model capable of diagnosing insulin resistance (HOMA-IR). As part of the study, a dataset was built using existing paediatric patient data containing subjects with and without insulin resistance. A novel machine learning model was then developed to predict the presence of insulin resistance based on dependent biometric variables with an optimal level of accuracy. This model is made publicly available through the implementation of a clinical decision support system (CDSS) prototype. The model classifies insulin resistant individuals with 81% accuracy and 75% of individuals without insulin resistance. This gives an overall accuracy of 78%. The user testing feedback for the CDSS is largely positive. Best practices were followed for building the model in accordance to those set out in previous studies. The biometric profile of insulin resistance represented in the model is likely to become better fitted to that of insulin resistance in the general population as more data are aggregated from sources. The infrastructure of the CDSS has also been built so that cross platform integration will be possible in future work.The current methods used by clinicians to identify insulin resistance in children are limited by invasive and clinically expensive blood testing. The benefits of this model would be to reduce the cost of clinical diagnosis and as a result, could also be used as a screening tool in the general childhood population.

The success of E-Learning system depends on the retrieval of relevant learning contents of learner. The best method to acquire learner needs is to construct an efficient learner profile which has to comply with the Semantic Web. Semantic Web relies heavily on formal ontologies to structure data.The proposed work suggests an approach to construct an efficient ontology based semantic learner profile by achieving the following objectives: First step is to collect static data using questionnaire and dynamic data using web log files. Second step is to preprocess weblog files to retrieve learner interest using semantic representation of WordNet and to retrieve learning style using decision tree classifier with significant rules. Third step is to construct an ontology using the retrieved data and to update ontology automatically using semantic similarity with WordNet. Finally an efficient fuzzy semantic retrieval is obtained using fuzzy linguistic variable which improves information retrieval and filtering.

The plasticity in our brain gives us promising ability to learn and know the world. Although great successes have been achieved in many fields, few bio-inspired methods have mimiced this ability. They are infeasible when the data is time-varying and the scale is large because they need all training data loaded into memory. Furthermore, even the popular deep convolutional neural network (CNN) models have relatively fixed structures. Through incremental PCANet, this paper aims at exploring a lifelong learning framework to achieve the plasticity of both feature and classifier constructions. The proposed model mainly comprises of three parts: Gabor filters followed by maxpooling layer offering shift and scale tolerance to input samples, cascade incremental PCA to achieve the plasticity of feature extraction and incremental SVM to pursue plasticity of classifier construction. Different from CNN, the plasticity in our model has no back propogation (BP) process and don’t need huge parameters. Experiments have been done and their results validate the plasticity of our models in both feature and classifier constructions and further verify the hypothesis of physiology that the plasticity of high layer is better than the low layer.

People need to know other people’s opinions to make well-informed decisions to buy products or services. Companies and organizations need to understand people’s attitude towards their products and services and use feedback from the customers to improve their products. Sentiment analysis techniques address these needs. While the majority of Internet users are not English speakers, most research papers in the sentiment-analysis field focus on English; resources for other languages are scarce. In this paper, we introduce a Persian sentiment lexicon, which consists of 1500 words along with their part-of-speech tags and polarity scores. We have used two machine-learning algorithms to evaluate the performance of this resource on a sentiment analysis task. The lexicon is freely available and can be downloaded from our website.

Image set representation and classification is an important problem in computer vision and pattern recognition area. In real application, image set data often come with kinds of noises, corruptions or large errors which usually make the recognition/learning tasks of image set more challengeable. In this paper, we utilize the low-rank representation/component of image set to represent the observed image set which is called Low-rank Image Set Representation (LRISR). Comparing with original observed image set, LRISR is generally noiseless and thus can encourage more robust learning process. Based on LRISR, we then use covariate-relation graph to encode the geometric relationship between covariates/features of LRISR and thus extract description vectors for LRISR classification task. Experimental results on several datasets demonstrate the benefits of the proposed image set representation and classification method.

The concept of using visual information as part of audio speech processing has been of significant recent interest. This paper presents a data driven approach that considers estimating audio speech acoustics using only temporal visual information without considering linguistic features such as phonemes and visemes. Audio (log filterbank) and visual (2D-DCT) features are extracted, and various configurations of MLP and datasets are used to identify optimal results, showing that given a sequence of prior visual frames an equivalent reasonably accurate audio frame estimation can be mapped.

Recurrent collaterals in the brain represent the recollection and execution of various monotonous activities such as breathing, brushing our teeth, chewing, walking, etc. These recurrent collaterals are found throughout the brain, each pertaining to a specific activity. Any deviation from regular activity falls back to the original cycle of activities, thus exhibiting a limit cycle or attractor dynamics. Upon analysis of some of these recurrent collaterals from different regions of the brain, it is observed that rhythmic theta oscillations play a vital role coordinating the functionalities of different regions of the brain. The neuromodulator acetylcholine, is found to be present in almost all of the regions where recurrent collaterals are present. This notable observation points to an underlying link between the generation and functioning of theta oscillations present in these recurrent collaterals, with the neuromodulator acetylcholine. Further, we show that these recurrent collaterals can be mathematically modeled using continuous time recurrent neural networks to account for the frequency of action potentials which follow the excitatory-inhibitory-excitatory (E-I-E) and inhibitory-excitatory-inhibitory (I-E-I) model. As a first case study, we present a detailed preliminary analysis of the CA3 region of the hippocampus, which is one of the most widely studied recurrent collaterals network in the brain, known to be responsible for storing and recalling episodic memories and also learning tasks. The recurrent collaterals present in this region are shown to follow an E-I-E pattern, which is analyzed using a mathematical model derived from continuous time recurrent neural networks, using inputs from a leaky integrate-and-fire neuronal model.

Community structure is one of the important features of complex brain network. Recently, major efforts have been made to investigate the non-overlapping community structure of brain network. However, an important fact is often ignored that the community structures of most real networks are overlapping. In this paper, we propose a novel method called sparse symmetric non-negative matrix factorization (ssNMF) to detect the overlapping community structure of the brain functional network, by adding a sparse constraint on the standard symmetric NMF (symNMF). Besides, we apply a sparse-network based framework by using non-negative adaptive sparse representation (NASR) to construct a sparse brain network. Simulated fMRI experimental results show that NMF-based methods achieve higher accuracy than methods of modularity optimization, normalized cuts and affinity propagation. Results of real fMRI experiments also lead to meaningful findings, which can help to promote the understanding of brain functional systems.