Neural Information Processing

To build a robust system for predicting emotions from user-generated videos is a challenging problem due to the diverse contents and the high level abstraction of human emotions. Evidenced by the recent success of deep learning (e.g. Convolutional Neural Networks, CNN) in several visual competitions, CNN is expected to be a possible solution to conquer certain challenges in human cognitive processing, such as emotion prediction. The emotion wheel (a widely used emotion categorization in psychology) may provide a guidance on building basic cognitive structure for CNN feature learning. In this work, we try to predict emotions from user-generated videos with the aid of emotion wheel guided CNN feature extractors. Experimental results show that the emotion wheel guided and CNN learned features improved the average emotion prediction accuracy rate to 54.2 %, which is better than that of the related state-of-the-art approaches.

The combination of modern reinforcement learning and deep learning approaches brings significant breakthroughs to a variety of domains requiring both rich perception of high-dimensional sensory inputs and policy selection. A recent significant breakthrough in using deep neural networks as function approximators, termed Deep Q-Networks (DQN), proves to be very powerful for solving problems approaching real-world complexities such as Atari 2600 games. To remove temporal correlation between the observed transitions, DQN uses a sampling mechanism called experience reply which simply replays transitions at random from the memory buffer. However, such a mechanism does not exploit the importance of transitions in the memory buffer. In this paper, we use prioritized sampling into DQN as an alternative. Our experimental results demonstrate that DQN with prioritized sampling achieves a better performance, in terms of both average score and learning rate on four Atari 2600 games.

This study proposes model-free deep inverse reinforcement learning to find nonlinear reward function structures. It is based on our previous method that exploits the fact that the log of the ratio between an optimal state transition and a baseline one is given by a part of reward and the difference of the value functions under linearly solvable Markov decision processes and reward and value functions are estimated by logistic regression. However, reward is assumed to be a linear function whose basis functions are prepared in advance. To overcome this limitation, we employ deep neural network frameworks to implement logistic regression. Simulation results show our method is comparable to model-based previous methods with less computing effort in the Objectworld benchmark. In addition, we show the optimal policy, which is trained with the shaping reward using the estimated reward and value functions, outperforms the policies that are used to collect data in the game of Reversi.

In this paper, we propose a novel learning method for the combined knowledge acquisition model. The combined knowledge acquisition model is a model for knowledge acquisition in which an agent heuristically find new knowledge by integrating existing plural knowledge. In the conventional model, there are two separate phases for combined knowledge acquisition: (a) solving a task with existing knowledge by trial and error and (b) learning new knowledge based on the experience in solving the task. However, since these two phases are carried out serially, the efficiency of learning was poor. In this paper, in order to improve this problem, we propose a novel knowledge acquisition method which realizes two phases simultaneously. Computer simulation results show that the proposed method much improves the efficiency of learning new knowledge.

Aiming for the emergence of higher functions such as “logical thinking”, our group has proposed completely novel reinforcement learning where exploration is performed based on the internal dynamics of a chaotic neural network. In this paper, in the learning of an obstacle avoidance task, it was examined that in the process of growing the dynamics through learning, the level of exploration changes from “lower” to “higher”, in other words, from “motor level” to “more abstract level”. It was shown that the agent learned to reach the goal while avoiding the obstacle and there is an area where the agent looks to pass through the right side or left side of the obstacle randomly. The result shows the possibility of the “higher exploration” though the agent sometimes collided with the obstacle and was trapped for a while as learning progressed.

PM₁₀ is one of contributors to air pollution. One cause of increases in PM₁₀ concentration in ambient air is the dust of bare land from rivers in drought season. The Taan and Tachia river are this study area, and data on PM₁₀ concentration, PM_2.5 concentration and meteorological condition at air monitoring site are used to establish a model for predicting next PM₁₀ concentration (PM₁₀(T + 1)) based on an artificial neural network (ANN) and to establish a mechanism for warning about PM₁₀(T + 1) concentration exceed 150 μg/m3 from rivers in drought season. The optimal architecture of an ANN for predicting PM₁₀(T + 1) concentration has six input factors include PM₁₀, PM_2.5 and meteorological condition. The train and test R was 0.8392 and 0.7900. PM₁₀(T) was the most important factor in predicting PM₁₀(T + 1) by sensitivity analysis. Finally, mechanism constraints were established for warning of high PM₁₀(T + 1) concentrations in river basins.

Currently, the World Wide Web (WWW) is the primary resource for cloud services information, including offers and providers. Cloud applications (Software as a Service), such as Google App, are one of the most popular and commonly used types of cloud services. Having access to a large amount of information on SaaS offers is critical for the potential cloud client to select and purchase an appropriate service. Web harvesting has become a primary tool for discovering knowledge from the Web source. This paper describes the design and development of Web scraper to collect information on SaaS offers from target Digital cloud services advertisement portals, namely www.getApp.com, and www.cloudreviews.com. The collected data were used to establish two datasets: a SaaS provider’s dataset and a SaaS reviews/feedback dataset. Further, we applied sentiment analysis on the reviews dataset to establish a third dataset called the SaaS sentiment polarity dataset. The significance of this study is that the first work focuses on Web harvesting for cloud computing domain, and it also establishes the first SaaS services datasets. Furthermore, we present statistical data that can be helpful to determine the current status of SaaS services and the number of services offered on the Web. In our conclusion, we provide further insight into improving Web scraping for SaaS service information. Our datasets are available online through www.bluepagesdataset.com.

Organizations need to continuously monitor, source and process large amount of operational data for optimizing the cloud computing environment. The research problem is: what are cloud monitoring data challenges – in particular virtual CPU monitoring data? This paper adopts a Systematic Literature Review (SLR) approach to identify and report cloud monitoring data challenges. SLR approach was applied to initially identify a large set of 1861 papers. Finally, 24 of 1861 relevant papers were selected and reviewed to identify the five major challenges of cloud monitoring data: monitoring technology, virtualization technology, energy, availability and performance. The results of this review are expected to help researchers and practitioners to understand cloud computing data challenges and develop innovative techniques and strategies to deal with these challenges.

Recommender systems provide personalized suggestions by learning users’ preference based on their historical feedback. To alleviate the heavy relying on historical data, several online recommendation methods are recently proposed and have shown the effectiveness in solving data sparsity and cold start problems in recommender systems. However, existing online recommendation methods neglect the use of social connections among users, which has been proven as an effective way to improve recommendation accuracy in offline settings. In this paper, we investigate how to leverage social connections to improve online recommendation performance. In particular, we formulate the online social recommendation task as a contextual bandit problem and propose a Locality-sensitive Linear Bandit (LS.Lin) method to solve it. The proposed model incorporates users’ local social relations into a linear contextual bandit model and is capable to deal with the dynamic changes of user preference and the network structure. We provide a theoretical analysis to the proposed LS.Lin method and then demonstrate its improved performance for online social recommendation in empirical studies compared with baseline methods.

In crowdsourcing systems, task recommendation can help workers to find their right tasks faster as well as help requesters to receive good quality output quicker. A number of previous works adopted active learning for task recommendation in crowdsourcing systems to achieve certain accuracy with a very low cost. However, the model updating methods in previous works are not suitable for real-world applications. In our paper, we propose a generic online-updating method for learning a factor analysis model, ActivePMF on TaskRec (Probabilistic Matrix Factorization with Active Learning on Task Recommendation Framework), for crowdsourcing systems. The larger the profile of a worker (or task) is, the less important is retraining its profile on each new work done. In case of the worker (or task) having large profile, our algorithm only retrains the whole feature vector of the worker (or task) and keeps all other entries in the matrix fixed. Besides, our algorithm runs batch update to further improve the performance. Experiment results show that our online-updating approach is accurate in approximating to a full retrain while the average runtime of model update for each work done is reduced by more than 90 % (from a few minutes to several seconds).

The medial temporal lobe is crucial for the encoding and retrieval of episodic long-term memories. It is widely assumed that memory encoding is associated with information transfer from sensory regions via the rhinal cortex into the hippocampus. Retrieval of information should then be associated with transfer in the reverse direction. However, experimental evidence for this mechanism is still lacking. Here, we show in human intracranial EEG data during two independent recognition memory paradigms that rhinal-hippocampal information flow significantly changes its directionality from encoding to retrieval. Using a novel phase-based method to analyze directional coupling of oscillations, coupling values were more positive (i.e., from rhinal cortex to the hippocampus) during encoding as compared to retrieval. These effects were observed in the delta (1–3 Hz) range where rhinal-hippocampal post-stimulus phase synchronization increased most robustly across both experiments.

Synaptic connections in local cortical circuit are highly heterogeneous and nonrandom. A few strong synaptic connections often form “cluster” that is a tightly connected group of several neurons. Global structure of the clusters, however, has not been clarified yet. It is unclear whether clusters distribute independently and isolated in cortical network, or these clusters are a part of large-scale of global network structure. Here, we develop a network model based on recent experimental data of V1. In addition to reproducing previous result of highly skewed EPSPs, the model also allows us to study mutual relationship and global feature of clusters. We find that the network consists with two largely different sub-networks; a small-world network consists only of a few strong EPSPs and a random network consists of dense weak EPSPs. In other words, local cortical circuit shows a duality, and previously reported clusters are results of local observation of the global small-world network.

EEG microstate of the brain has been suggested to reflect functional significance of cognitive activity. In this paper, from math-gifted and non-gifted adolescents’ EEG during a reasoning task, four classes of microstate configuration were extracted based on clustering analysis approach. Computations of multiple parameters were down for each class of EEG microstate. Between-groups statistical and discriminating analyses for these parameters discovered significant functional differences between math-gifted and non-gifted subjects in momentary microstates, involving mean duration and occurrence of EEG electric field configuration. Additionally, the topological differences between the two groups vary across classes and reflect functional disassociation of cognitive processing of the reasoning task. Our study suggests that the microstate classes can be used as the effective EEG features for identifying mental operations by individuals with typical cognitive ability differences.

Figure-ground (FG) segregation is a crucial function of the intermediate-level vision. Physiological studies on monkey V2 have reported border-ownership (BO) selective cells that signal the direction of figure along a local border. However, local borders in natural images are often complicated and they often do not provide a clue for FG segregation. In the present study, we hypothesize that a population of V4 cells represents FG by means of surface rather than border. We investigated this hypothesis by the computational analysis of neural signals from multiple cells in monkey V4. Specifically, we applied Support Vector Machine as an ideal integrator to the cellular responses, and examined whether the responses carry information capable of determining correct local FG. Our results showed that the responses from several tens of cells are capable of determining correct local FG in a variety of natural image patches while single-cell responses hardly determine FG, suggesting a population coding of local FG by a small number of cells in V4.

Electroencephalography (EEG) based preliminary examination has been widely used in diagnosis of brain diseases. Based on previous studies, clinical brain death determination also can be actualized by analyzing EEG signal of patients. Dynamic Multivariate empirical mode decomposition (D-MEMD) and approximate entropy (ApEn) are two kinds of methods to analyze brain activity status of the patients in different perspectives for brain death determination. In our previous studies, D-MEMD and ApEn methods were always used severally and it cannot analyzing the patients’ brain activity entirety. In this paper, we present a combine analysis method based on D-MEMD and ApEn methods to determine patients’ brain activity level. Moreover, We will analysis three different status EEG data of subjects in normal awake, comatose patients and brain death. The analyzed results illustrate the effectiveness and reliability of the proposed methods.

In this paper, we develop an event-based adaptive robust stabilization method for continuous-time nonlinear systems with uncertain terms via a self-learning technique called neural dynamic programming. Through system transformation, it is proven that the robustness of the uncertain system can be achieved by designing an event-triggered optimal controller with respect to the nominal system under a suitable triggering condition. Then, the idea of neural dynamic programming is adopted to perform the main controller design task by building and training a critic network. Finally, the effectiveness of the present adaptive robust control strategy is illustrated via a simulation example.

This paper analyzes entropy of occupancy grid map (OGM) for evaluating registration performance of SLAM (simultaneous localization and mapping) algorithms. So far, there are a number of SLAM algorithms having been proposed, but we do not have general measure to evaluate the registration performance of point clouds obtained by LRF (laser range finder) for SLAM algorithms. This paper analyzes to show that good registration seems corresponding to large overlap of point clouds in OGM as well as large entropy, large uncertainty and low information of OGM. This analysis indicates a method of entropy maximization of OGM for selecting good registration of SLAM algorithms. By means of executing numerical experiments, we show the validity and the effectiveness of the entropy of OGM to evaluate the registration performance.

A service robot requires natural and interactive interaction with users without explicit commands. It is still one of the difficult problems to generate robust reactions for the robot in the real environment with unreliable sensor data to satisfy user’s requests. This paper presents an intention-response model based on mirror neuron and theory of mind, and analyzes the performance for a humanoid to show the usefulness. The model utilizes the modules of behavior selection networks to realize prompt response and goal-oriented characteristics of the mirror neuron, and performs reactions according to an action plan based on theory of mind. To cope with conflicting goals, behaviors of the sub-goal unit are generated using a hierarchical task network. Experiments with various scenarios reveal that appropriate reactions are generated according to external stimuli.

In this paper, a novel dynamic surface sliding mode control method is proposed for three-dimensional trajectory tracking control of autonomous underwater vehicle (AUV) in the presence of model errors. To enhance the robustness, the sliding mode control approach is modified by employing dynamic surface control (DSC). The radial basis function neural network (RBFNN) approximation technique is used for approximating model errors, furthermore the norm of the ideal weighting vector in neural network system is considered as the estimation parameter, such that only one parameter is adjusted. The proposed controller guarantees uniform ultimate boundedness (UUB) of all the signals in the closed-loop system via Lyapunov stability analysis, while the tracking errors converge to a small neighborhood of the desired trajectory. Finally, simulation studies are given to illustrate the performance of the proposed algorithm.

Reservoir computing is a computational framework suited for sequential data processing, consisting of a reservoir part and a readout part. Not only theoretical and numerical studies on reservoir computing but also its implementation with physical devices have attracted much attention. In most studies, the reservoir part is constructed with identical units. However, a variability of physical units is inevitable, particularly when implemented with nano/micro devices. Here we numerically examine the effect of variability of reservoir units on computational performance. We show that the heterogeneity in reservoir units can be beneficial in reducing the prediction error in the reservoir computing system with a simple cycle reservoir.

How can we build robot controllers that are able to work under harsh conditions, but without experiencing catastrophic failures? As seen on the recent Fukushima’s nuclear disaster, standard robots break down when exposed to high radiation environments. Here we present the results from two arrangements of Spiking Neural Networks, based on the Liquid State Machine (LSM) framework, that were able to gracefully degrade under the effects of a noisy current injected directly into each simulated neuron. These noisy currents could be seen, in a simplified way, as the consequences of exposition to non-destructive radiation. The results show that not only can the systems withstand noise, but one of the configurations, the Modular Parallel LSM, actually improved its results, in a certain range, when the noise levels were increased. Also, the robot controllers implemented in this work are suitable to run on a modern, power efficient neuromorphic hardware such as SpiNNaker.

We investigate reservoir computing systems whose dynamics are at critical bifurcation points based on center manifold theorem. We take echo state networks as an example and show that the center manifold defines mapping of the input dynamics to higher dimensional space. We also show that the mapping by center manifolds can contribute to recognition of attractors of input dynamics. The implications for realization of reservoir computing as real physical systems are also discussed.

This paper proposes a new approach to object tracking using the Hybrid Gravitational Search Algorithm (HGSA). HGSA introduces the Gravitational Search Algorithm (GSA) to the field of object tracking by incorporating Particle Swarm Optimization (PSO) using a novel weight function that elegantly combines GSA’s gravitational update component with the cognitive and social components of PSO. The hybridized algorithm acquires PSO’s exploitation of past information and fast convergence property while retaining GSA’s capability in fully utilizing all current information. The proposed framework is compared against standard natural phenomena based algorithms and Particle Filter. Experiment results show that HGSA largely reduces convergence to local optimum and significantly out-performed the standard PSO algorithm, the standard GSA and Particle Filter in terms of tracking accuracy and stability under occlusion and non-linear movement in a large search space.

Reservoir computing is a novel paradigm of neural network, offering advantages in low learning cost and ease of implementation as hardware. In this paper we propose a concept of reservoir computing consisting of a semiconductor laser subject to external feedback by a mirror, where input signal is supplied as modulation pattern of mirror reflectivity. In that system, non-linear interaction between optical field and electrons are enhanced in complex manner under substantial external feedback, leading to achieve highly nonlinear projection of input electric signal to output optical field intensity. It is exhibited that the system can most efficiently classify waveforms of sequential input data when operating around laser oscillation’s effective threshold.

Deep neural networks (DNNs) are a state-of-the-art processing model in the field of machine learning. Implementation of DNNs into embedded systems is required to realize artificial intelligence on robots and automobiles. Embedded systems demand great processing speed and low power consumption, and DNNs require considerable processing resources. A field-programmable gate array (FPGA) is one of the most suitable devices for embedded systems because of their low power consumption, high speed processing, and reconfigurability. Autoencoders (AEs) are key parts of DNNs and comprise an input, a hidden, and an output layer. In this paper, we propose a novel hardware implementation of AEs having shared synapse architecture. In the proposed architecture, the value of each weight is shared in two interlayers between input-hidden layer and hidden-output layer. This architecture saves the limited resources of an FPGA, allowing a reduction of the synapse modules by half. Experimental results show that the proposed design can reconstruct input data and be stacked. Compared with the related works, the proposed design is register transfer level description, synthesizable, and estimated to decrease total processing time.

Time-domain weighted-sum operation based on a spiking neuron model is discussed and evaluated from a VLSI implementation point of view. This calculation model is useful for extremely low-power operation because transition states in resistance and capacitance (RC) circuits can be used. Weighted summation is achieved with energy dissipation on the order of 1 fJ using the current CMOS VLSI technology if 1 G$$\varOmega $$Ω order resistance can be used, where the number of inputs can be more than a hundred. This amount of energy is several orders of magnitude lower than that in conventional digital processors. In this paper, we show the software simulation results that verify the proposed calculation method for a 500-input neuron in a three-layer perceptron for digit character recognition.

Boltzmann machines are a useful model for deep neural networks in artificial intelligence, but in their software or hardware implementation, they require random number generation for stochastic operation, which consumes considerable computational resources and power. Chaotic Boltzmann machines (CBMs) have been proposed as a model using chaotic dynamics instead of stochastic operation. They require no random number generation, and are suitable for analog VLSI implementation. In this paper, we describe software simulation results for CBM operation, and propose a CMOS circuit of CBMs using the subthreshold operation of MOSFETs.

We have been developing the “Neurocommunicator”, an EEG-based communication aid for people with severe motor disabilities. This system analyzes an event-related potential (ERP) to the sequentially flashed pictograms to indicate a desired message, and predicts the user’s choice in the brain. To speed-up of this decoding process, we introduced a special algorithm, the Virtual Decision Function (VDF), which was originally designed to reflect the continuous progress of binary decisions on a single trial basis of neuronal activities in the primate brain. We applied the VDF to the EEG signals, and succeeded in faster decoding of the target.

The echo state network is a framework for temporal data processing, such as recognition, identification, classification and prediction. The echo state network generates spatiotemporal dynamics reflecting the history of an input sequence in the dynamical reservoir and constructs mapping from the input sequence to the output one in the readout. In the conventional dynamical reservoir consisting of sparsely connected neuron units, more neurons are required to create more time delay. In this study, we introduce the dynamic synapses into the dynamical reservoir for controlling the nonlinearity and the time constant. We apply the echo state network with dynamic synapses to several benchmark tasks. The results show that the dynamic synapses are effective for improving the performance in time series prediction tasks.

In recent years, a breakthrough has been made in infant level AI due to the acquisition of representation, which was realized by deep learning. By this, the construction of AI that specializes in a specific task that does not require a high-level understanding of language is becoming a possibility. The primary remaining issue for the realization of human-level AI is the realization of general intelligence capable of solving flexible problems by combining highly reusable knowledge. Therefore, this research paper explores the possibility of approaching artificial general intelligence with such abilities based on mesoscopic connectome.

Risk-aware control is a new type of robust nonlinear stochastic controller in which state variables are represented by time-varying probability densities and the desired trajectory is replaced by a cost function that specifies both the goals of movement and the potential risks associated with deviations. Efficient implementation is possible using the theory of Stochastic Dynamic Operators (SDO), because for most physical systems the SDO operators are near-diagonal and can thus be implemented using distributed computation. I show such an implementation using 4.3 million spiking neurons simulated in real-time on a GPU. I demonstrate successful control of a commercial desktop robot for a visually-guided reaching task, and I show that the operators can be learned during repetitive practice using a recursive learning rule.

We describe a method of regularization for the restricted Bayesian network BESOM, which possesses a network structure similar to that of Deep Learning. Two types of penalties are introduced to avoid overfitting and local minimum problems. The win-rate penalty ensures that each value in the nodes is used evenly; the lateral-inhibition penalty ensures that the nodes in the same layer are independent. Bayesian networks with these prior distributions can be converted into equivalent Bayesian networks without prior distributions, then the EM algorithm becomes easy to be executed.

Whole brain architecture (WBA) which uses neural networks to imitate a human brain is attracting increased attention as a promising way to achieve artificial general intelligence, and distributed vector representations of words is becoming recognized as the best way to connect neural networks and knowledge. Distributed representations of words have played a wide range of roles in natural language processing, and they have become increasingly important because of their ability to capture a large amount of syntactic and lexical meanings or relationships. Relation vectors are used to represent relations between words, but this approach has some problems; some relations cannot be easily defined, for example, sibling relations, parent-child relations, and many-to-one relations. To deal with these problems, we have created a novel way of representing relations: we represent relations by planes instead of by vectors, and this increases by more than 10 % the accuracy of predicting the relation.

Emotion is a very popular but not well-known phenomenon of animals. Human emotion/feeling is more complex including the emotion features and the intelligent features. Though there are many researches on emotion/feeling, its computational role on self-maintenance is not known well. But it must be important because most of animals look to have similar emotion and there must be a reason for its similarity. Therefore, in this paper, we discuss on a possible component of emotion system, compare their computational model, and propose a possible hypothesis that the emotion is a system of value calculation for a decision making. For a discussion, we show a possible computational model of feeling system in brain.

The Whole Brain Architecture Initiative is a non-profit organization (NPO) founded in Japan in August 2015, whose purpose is to support research activities aiming for realizing artificial intelligence with human-like cognitive capabilities by studying the entire architecture of the brain. It performs educational activities such as holding seminars and hackathons and compiling educational materials, as well as R&D activities such as developing software platforms to support research in artificial intelligence and facilitating communication among research communities.

We consider a system of two-level quantum quasi-spins and gauge bosons put on a 3+1D lattice. As a model of neural network of the brain functions, these spins describe neurons quantum-mechanically, and the gauge bosons describes weights of synaptic connections. It is a generalization of the Hopfield model to a quantum network with dynamical synaptic weights. At the microscopic level, this system becomes a model of quantum brain dynamics proposed by Umezawa et al., where spins and gauge field describe water molecules and photons, respectively. We calculate the phase diagram of this system under quantum and thermal fluctuations, and find that there are three phases; confinement, Coulomb, and Higgs phases. Each phase is classified according to the ability to learn patterns and recall them. By comparing the phase diagram with that of classical networks, we discuss the effect of quantum fluctuations and thermal fluctuations (noises in signal propagations) on the brain functions.

Brain-inspired Computing Architecture (BriCA) is a generic software platform for modular composition of machine learning algorithms. It can combine and schedule an arbitrary number of machine learning components in a brain-inspired fashion to construct higher level structures such as cognitive architectures. We would like to report and discuss the core concepts of BriCA version 1 and prospects toward future development.

Cognition, judgment, action, and expression acquisition have been widely treated in studies on recently developed deep learning. However, although each study has been specialised for specific tasks and goals, cognitive architecture that integrates many different functions remains necessary for the realisation of artificial general intelligence. To that end, a cognitive architecture fully described with restricted Boltzmann machines (RBMs) in a unified way are promising, and we have begun to implement various cognitive functions with an RBM base. In this paper, we propose new stacked half RBMs (SHRBMs) made from layered half RBMs (HRBMs) that handle working memory. We show that an ability to solve maze problems that requires working memory improves drastically when SHRBMs in the agent’s judgment area are used instead of HRBMs or other RBM-based models.

Artificial General Intelligence (AGI) refers to machine intelligence that can effectively conduct variety of human tasks. Therefore AGI research requires multivariate and realistic learning environments. In recent years, game engines capable of constructing highly realistic 3D virtual worlds have also become available at low cost. In accordance with these changes, we developed the “Life in Silico” (LIS) framework, which provides virtual agents with learning algorithms and their learning environments with game engine. This should in turn allow for easier and more flexible AGI research. Furthermore, non-experts will be able to play with the framework, which would enable them to research as their hobby. If AGI research becomes popular in this manner, we may see a sudden acceleration towards the “Democratization of AGI”.

The mean firing rate of a border-ownership selective (BOS) neuron encodes where a foreground figure relative to its classical receptive field. Physiological experiments have demonstrated that top-down attention increases firing rates and decreases spike synchrony between them. To elucidate mechanisms of attentional modulation on rates and synchrony of BOS neurons, we developed a spiking neuron network model: BOS neurons receive synaptic input which reflects visual input. The synaptic input strength is modulated multiplicatively by the activity of Grouping neurons whose activity represents the object’s location and mediates top-down attentional projection to BOS neurons. Model simulations agree with experimental findings, showing that attention to an object increases the firing rates of BOS neurons representing it while decreasing spike synchrony between pairs of such neurons. Our results suggest that top-down attention multiplicatively emphasizes synaptic current due to bottom-up visual inputs.

One of the traditional models for finding the location of a mobile source is the time-of-arrival (TOA). It usually assumes that the measurement noise follow a Gaussian distribution. However, in practical, outliers are difficult to be avoided. This paper proposes an $$l_1$$l1-norm based objective function for alleviating the influence of outliers. Afterwards, we utilize the Lagrange programming neural network (LPNN) framework for the position estimation. As the framework requires that its objective function and constraints should be twice differentiable, we introduce an approximation for the $$l_1$$l1-norm term in our LPNN formulation. From the simulation result, our proposed algorithm has very good robustness.

We have expected that dynamic higher functions such as “thinking” emerge through the growth from exploration in the framework of reinforcement learning (RL) using a chaotic Neural Network (NN). In this frame, the chaotic internal dynamics is used for exploration and that eliminates the necessity of giving external exploration noises. A special RL method for this framework has been proposed in which “traces” were introduced. On the other hand, reservoir computing has shown its excellent ability in learning dynamic patterns. Hoerzer et al. showed that the learning can be done by giving rewards and exploration noises instead of explicit teacher signals. In this paper, aiming to introduce the learning ability into our new RL framework, it was shown that the memory-required task in the work of Hoerzer et al. could be learned without giving exploration noises by utilizing the chaotic internal dynamics while the exploration level was adjusted flexibly and autonomously. The task could be learned also using “traces”, but still with problems.

Directional traveling waves are organized in a hippocampal CA3 recurrent network model composed of biophysical pyramidal cells and inhibitory interneurons with gap junctions. The network spontaneously organizes neuronal activities traveling in a particular direction and the organized traveling waves are modified by repetitive local inputs. We found that the distributions of inter-spike intervals (ISIs) of pyramidal cells and interneurons are involved with spontaneous traveling waves that can be modified by local stimulation. Similar ISI distributions emerge in a network that has no gap junctions, but strong mutual connections between pyramidal cells and interneurons. These results suggest that interaction between interneurons through gap junctions contributes to enhancing the inhibition of pyramidal cells for organizing traveling waves.

The current paper presents how performance of a particular deterministic dynamical neural network model in predictive coding scheme differ when it is trained for a set of prototypical movement patterns using their modulated teaching samples from when it is trained using unmodulated teaching samples. Multiple timescale neural network (MTRNN) trained with or without modulated patterns was applied in a simple numerical experiment for a task of imitative synchronization by inferencing the internal states by the error regression, and the results suggest that the scheme of training with modulated patterns can outperform the scheme of training without them. In our second experiment, our network was tested with naturally fluctuated movement patterns in an imitative interaction between a robot and different human subjects, and the results showed that a network trained with fluctuated patterns could achieve generalization in learning, and mutual imitation by synchronization was obtained.

Graph Cuts has became increasingly useful methods for the image segmentation. In Graph Cuts, given images are replaced by grid graphs, and the image segmentation process is performed using the minimum cut (min-cut) algorithm on the graphs. For Graph Cuts, the most typical min-cut algorithm is the B-K algorithm. While the B-K algorithm is very efficient, it is still far from real-time processing. In addition, the B-K algorithm gives only the single min-cut even if the graph has multiple-min-cuts. The conventional Graph Cuts has a possibility that a better minimum cut for an image segmentation is frequently overlooked. Therefore, it is important to apply a more effective min-cut algorithm to Graph Cuts. In this research, we propose a new image segmentation technique using Graph Cuts based on the maximum-flow neural network (MF-NN). The MF-NN is our proposed min-cut algorithm based on a nonlinear resistive circuit analysis. By applying the MF-NN to Graph Cuts instead of the B-K algorithm, image segmentation problems can be solved as the nonlinear resistive circuits analysis. In addition, the MF-NN has an unique feature that multiple-min-cuts can be find easily. That is, it can be expected that our proposed method can obtain more accurate results than the conventional Graph Cuts which generates only one min-cut. When the proposed circuit model is designed with the integrated circuit which can change graph structure and branch conductance, a novel image segmentation technique with real-time processing can be expected.

This paper considers the joint routing and bitrate adjustment optimization for DASH (Dynamic Adaptive Streaming over HTTP) video service using neuro-dynamic programming (NDP) in software-defined networking (SDN). We design an open optimization architecture based on OpenFlow based SDN. Following this architecture, we formulate the joint routing and bitrate adjustment problem as a Markov Decision Process (MDP) for maximizing the average reward. In order to solve the curses of dimensionality, we employ neuro-dynamic programming method to conceive an online learning framework and develop a NDP based joint routing and bitrate adjustment algorithm for DASH video service. At last, an emulation platform based on POX and Mininet is constructed to verify the performance of the proposed algorithm. The experimental results indicate our algorithm has more excellent performance compared with OSPF based algorithm.

This paper studies dynamic binary neural networks that can generate various periodic orbits. The networks is characterized by signum activation function and ternary connection parameters. In order to analyze the dynamics, we present two simple feature quantities that characterize plentifulness of transient phenomena and superstability of the periodic orbits. Calculating the feature quantities for a class of networks, we investigate transient and superstability of the periodic orbits.

Fluctuating activities in the deterministic chaos cause a phenomenon that is similar to stochastic resonance (SR) whereby the presence of noise helps a non-linear system to amplify a weak (under-barrier) signal. In this phenomenon, called chaotic resonance (CR), the system responds to the weak input signal by the effect of intrinsic chaotic activities under the condition where no additive noise exists. Recently, we have revealed that the signal response of the CR in the spiking neuron model has an unimodal maximum with respect to the degree of stability for chaotic orbits quantified by maximum Lyapunov exponent. In response to this situation, in this study, focusing on CR in the systems with chaos-chaos intermittency, we examine the signal response in a cubic map and a chaotic neural network embedded two symmetric patterns by cross correlation and Lyapunov exponent (or maximum Lyapunov exponent). As the results, it is confirmed that the efficiency of the signal response has a peak at the appropriate instability of chaotic orbit in both systems. That is, the instability of chaotic orbits in CR can play a role the noise strength of SR in not only spiking neural systems but also the systems with chaos-chaos intermittency.

Extreme Learning Machine (ELM) is a universal approximation method that is extremely fast and easy to implement, but the weights of the model are normally randomly selected so they can lead to poor prediction performance. In this work, we applied Weighted Similarity Extreme Learning Machine in combination with Jaccard/Tanimoto (WELM-JT) and cluster analysis (namely, k-means clustering and Support Vector Clustering) on similarity and distance measures (i.e., Jaccard/Tanimoto and Euclidean) in order to predict which compounds with not-so-different chemical structures have an activity for treating a certain symptom or disease. The proposed method was experimented on one of the most challenging datasets named Maximum Unbiased Validation (MUV) dataset with 4 different types of fingerprints (i.e. ECFP_4, ECFP_6, FCFP_4 and FCFP_6). The experimental results show that WELM-JT in combination with k-means-ED gave the best performance. It retrieved the highest number of active molecules and used the lowest number of nodes. Meanwhile, WELM-JT with k-means-JT and ECFP_6 encoding proved to be a robust contender for most of the activity classes.

Since labor intensive and time consuming issue, manual curation in metabolic information extraction currently was replaced by text mining (TM). While TM in metabolic domain has been attempted previously, it is still challenging due to variety of specific terms and their meanings in different contexts. Named Entity Recognition (NER) generally used to identify interested keyword (protein and metabolite terms) in sentence, this preliminary task therefore highly influences the performance of metabolic TM framework. Conditional Random Fields (CRFs) NER has been actively used during a last decade, because it explicitly outperforms other approaches. However, an efficient CRFs-based NER depends purely on a quality of corpus which is a nontrivial task to produce. This paper introduced a hybrid solution which combines CRFs-based NER, dictionary usage, and complementary modules (constructed from existing corpus) in order to improve the performance of metabolic NER and another similar domain.

Revealing the underlying complex architecture of human diseases has received considerable attention since the exploration of genotype-phenotype relationships in genetic epidemiology. Identification of these relationships becomes more challenging due to multiple factors acting together or independently. A deep neural network was trained in the previous work to identify two-locus interacting single nucleotide polymorphisms (SNPs) related to a complex disease. The model was assessed for all two-locus combinations under various simulated scenarios. The results showed significant improvements in predicting SNP-SNP interactions over the existing conventional machine learning techniques. Furthermore, the findings are confirmed on a published dataset. However, the performance of the proposed method in the higher-order interactions was unknown. The objective of this study is to validate the model for the higher-order interactions in high-dimensional data. The proposed method is further extended for unsupervised learning. A number of experiments were performed on the simulated datasets under same scenarios as well as a real dataset to show the performance of the extended model. On an average, the results illustrate improved performance over the previous methods. The model is further evaluated on a sporadic breast cancer dataset to identify higher-order interactions between SNPs. The results rank top 20 higher-order SNP interactions responsible for sporadic breast cancer.

Treebanks for clinical text are not enough for supervised dependency parsing no matter in their scale or diversity, leading to still unsatisfactory performance. Many unlabeled text from web can make up for the scarceness of treebanks in some extent. In this paper, we propose to gain syntactic knowledge from web text as syntactic cluster features to improve dependency parsing on clinical text. We parse the web text and compute the distributed representation of each words base on their contexts in dependency trees. Then we cluster words according to their distributed representation, and use these syntactic cluster features to solve the data sparseness problem. Experiments on Genia show that syntactic cluster features improve the LAS (Labled Attachment Score) of dependency parser on clinical text by 1.62 %. And when we use syntactic clusters combining with brown clusters, the performance gains by 1.93 % on LAS.

Inferring gene regulatory networks (GRN) from microarray gene expression data is a highly challenging problem in computational and systems biology. To make GRN reconstruction process more accurate and faster, in this paper, we develop a technique to identify the gene having maximum in-degree in the network using the temporal correlation of gene expression profiles. The in-degree of the identified gene is estimated applying evolutionary optimization algorithm on a decoupled S-system GRN model. The value of in-degree thus obtained is set as the maximum in-degree for inference of the regulations in other genes. The simulations are carried out on in silico networks of small and medium sizes. The results show that both the prediction accuracy in terms of well known performance metrics and the computational time of the optimization process have been improved when compared with the traditional S-system model based inference.

In this paper, a sleep stage prediction method using respiration and body-movement based on probabilistic classifier is proposed. A pressure sensor is employed to capture respiratory signal. We propose to use least-squares probabilistic classifier (LSPC), a computationally effective probabilistic classifier, for four-class sleep stage classification (wakefulness, rapid-eye movement sleep, light sleep, deep sleep). Thanks to output of posterior probability of each class by LSPC, we can directly handle the confidence of predicted sleep stages. In addition, we introduce a method to handle imbalanced data problem which arises in sleep data collection. The experimental results demonstrate the effectiveness of sleep stage prediction by LSPC.

Removing ring artifacts in Cone Beam Computed Tomography (CBCT) images without impairing the image quality is critical for the application of CBCT. In this paper, we propose a novel method for the removal of ring artifacts in CBCT Images using an image smoothing based on relative total variation (RTV). After transforming the CBCT image into polar coordinates, we introduce a single-direction smoothing to separate the small scale textures, which include the artifacts, from the image structures. Then the artifact template is generated by median value extraction. Finally, the artifact template is transformed back into Cartesian coordinates and is subtracted from the original CBCT image. Experiments on different CBCT images show that the proposed method can obtain satisfactory results.

Heartbeat signal detection and/or monitoring is very important in the rescue of human beings existing under debris after disasters such as earthquakes as well as in the monitoring of patients in hospital. In this paper, we propose a human heartbeat detection/monitoring system employing chirp Z-transform and a time-sequential prediction neural network. The system is an adaptive radar using 2.5 GHz continuous microwave. The CZT realizes high resolution peak search in the frequency domain. We use a neural network to track adaptively the heartbeat signal which often has frequency fluctuation. The network learns the time-sequential peak frequency online in parallel to the detection and tracking. Even when the heartbeat frequency drifts, the network finds and tracks the heartbeat. Experiments demonstrate that the proposed system has high effectiveness in distinction between person-exist and person-non-exist observations, resulting in successful detection of persons.

We presented new reconstruction algorithms for compressed sensing magnetic resonance imaging (CS-MRI) based on the combination of the fast composite splitting algorithm (FCSA) and complex dual-tree wavelet transform (DT-CWT) and on the combination of FCSA and double density dual-tree wavelet transform (DDDT-DWT), respectively. We applied the bivariate thresholding to these two combinations. The proposed methods not only inherit the effectiveness and fast convergence of FCSA but also improve the sparse representation of both point-like and curve-like features. Experimental results validate the effectiveness and efficiency of the proposed methods.

In this paper we aim for the replication of a state of the art architecture for recognition of human actions using skeleton poses obtained from a depth sensor. We review the usefulness of accurate human action recognition in the field of robotic elderly care, focusing on fall detection. We attempt fall recognition using a chained Growing When Required neural gas classifier that is fed only skeleton joints data. We test this architecture against Recurrent SOMs (RSOMs) to classify the TST Fall detection database ver. 2, a specialised dataset for fall sequences. We also introduce a simplified mathematical model of falls for easier and faster bench-testing of classification algorithms for fall detection.The outcome of classifying falls from our mathematical model was successful with an accuracy of $$ 97.12 \pm 1.65\,\%$$97.12±1.65% and from the TST Fall detection database ver. 2 with an accuracy of $$90.2 \pm 2.68\,\%$$90.2±2.68% when a filter was added.

A botnet detection method using the graphical lasso is studied. Hamasaki et al. proposed a botnet detection method based on graphical lasso applied on darknet traffic, which captures change points of outputs of graphical lasso caused by a botnet activity. In their method, they estimate cooperative relationship of bots using graphical lasso. If the regularization coefficient of graphical lasso is appropriately tuned, it can remove false cooperative relationships to some extent. Though they represent the cooperative relationships of bots as a graph, they didn’t use its graphical properties. We propose a new method of botnet detection based on ‘graph density’, for which we introduce a new method to set the regularization coefficient automatically. The effectiveness of the proposed method is illustrated by experiments on darknet data.

The number of security incidents faced by Android users is growing, along with the surge in malware targeting Android terminals. Such malware arrives at the Android terminals in the form of Android Packages (APKs). Assorted techniques for protecting Android users from such malware have been reported, but most of them focus on the APK files themselves. Different from these approaches, we use metadata, such as web information obtained from the online APK markets, to improve the accuracy of malware identification. In this paper, we introduce malware detection schemes using metadata, which includes categories and descriptions of APKs. We introduce two types of schemes: statistical scheme and support vector machine-based scheme. Finally, we analyze and discuss the performance and usability of the schemes, and confirm the usability of web information for the purpose of identifying malware.

Reinforcement learning (RL) is one of the artificial intelligence approaches that has been deployed effectively to improve performance of distributed cognitive radio networks (DCRNs). However, in existing proposals that involve multi-agents, perceptions of the agents are shared in plain in order to calculate optimal actions. This raises privacy concern where an agent learns private information (e.g. Q-values) of the others, which can then be used to infer, for instance, the actions of these other agents. In this paper, we provide a preliminary investigation and a privacy-preserving protocol on multi-agent RL in DCRNs. The proposed protocol provides RL computations without revealing agents’ private information. We also discuss the security and performance of the protocol.

Wireless LAN (WLAN) service has been provided in many companies, universities, hotels, coffee shops and even on the street, which supports the growing number of users with mobile devices. Kyoto Women’s University has offered WLAN service with many access points in centralized control style since 2011, which allows mobile users to access the network at any location covered by its access points while on campus. It is useful for evacuation planning to figure out when and where people gather, and therefore, it is worth understanding the trends of WLAN usage in each organization at all times. In this paper, we analysed the trends of WLAN usage in the university and described some applications.

A method for botnet detection from traffic data of the Internet by the Non-negative Matrix Factorization (NMF) was proposed by (Yamauchi et al. 2012). This method assumes that traffic data is composed by several types of communications, and estimates the number of types in the data by the minimum description length (MDL) criterion. However, consideration on the MDL criterion was not sufficient and validity has not been guaranteed. In this paper, we refine the MDL criterion for NMF and report results of experiments for the new MDL criterion on synthetic and real data.

Spin-glass theory developed in statistical mechanics has found its usage in various information science problems. In this study, we focus on the application of spin-glass models in unsupervised and semi-supervised learning. Several key papers in this field are reviewed, to answer the question that why and how spin-glass is adopted. The question can be answered from two aspects.Firstly, adopting spin-glass models enables the vast knowledge base developed in statistical mechanics to be used, such as the self-organizing grains at the superparamagnetic phase has a natural connection to clustering. Secondly, spin-glass model can serve as a bridge for model development, i.e., one can map existing model into spin-glass manner, facilitate it with new features and finally map it back.

Non-negative Matrix Factorization (NMF) has been widely exploited to learn latent features from data. However, previous NMF models often assume a fixed number of features, say p features, where p is simply searched by experiments. Moreover, it is even difficult to learn binary features, since binary matrix involves more challenging optimization problems. In this paper, we propose a new Bayesian model called infinite non-negative binary matrix tri-factorizations model (iNBMT), capable of learning automatically the latent binary features as well as feature number based on Indian Buffet Process (IBP). Moreover, iNBMT engages a tri-factorization process that decomposes a nonnegative matrix into the product of three components including two binary matrices and a non-negative real matrix. Compared with traditional bi-factorization, the tri-factorization can better reveal the latent structures among items (samples) and attributes (features). Specifically, we impose an IBP prior on the two infinite binary matrices while a truncated Gaussian distribution is assumed on the weight matrix. To optimize the model, we develop an efficient modified maximization-expectation algorithm (ME-algorithm), with the iteration complexity one order lower than another recently-proposed Maximization-Expectation-IBP model [9]. We present the model definition, detail the optimization, and finally conduct a series of experiments. Experimental results demonstrate that our proposed iNBMT model significantly outperforms the other comparison algorithms in both synthetic and real data.

In this paper, we propose a novel manifold regularized online semi-supervised learning (OS$$^2$$2L) model in an Reproducing Kernel Hilbert Space (RK-HS). The proposed algorithm, named Model-Based Online Manifold Regularization (MOMR), is derived by solving a constrained optimization problem, which is different from the stochastic gradient algorithm used for solving the online version of the primal problem of Laplacian support vector machine (LapSVM). Taking advantage of the convex property of the proposed model, an exact solution can be obtained iteratively by solving its Lagrange dual problem. Furthermore, a buffering strategy is introduced to improve the computational efficiency of the algorithm. Finally, the proposed algorithm is experimentally shown to have a comparable performance to the standard batch manifold regularization algorithm.

Neural Networks (NN) have achieved great success in pattern recognition and machine learning. However, the success of NNs usually relies on a sufficiently large number of samples. When fed with limited data, NN’s performance may be degraded significantly. In this paper, we introduce a novel neural network called Memory Network, which can learn better from limited data. Taking advantages of the memory from previous samples, the new model could achieve remarkable performance improvement on limited data. We demonstrate the memory network in Multi-Layer Perceptron (MLP). However, it keeps straightforward to extend our idea to other neural networks, e.g., Convolutional Neural Networks (CNN). We detail the network structure, present the training algorithm, and conduct a series of experiments to validate the proposed framework. Experimental results show that our model outperforms the traditional MLP and other competitive algorithms in two real data sets.

Reinforcement learning aims at solving stochastic sequential decision making problems through direct trial-and-error interactions with the learning environment. In this paper, we will develop generalized compatible features to approximate value functions for reliable Reinforcement Learning. Further guided by an Actor-Critic Reinforcement Learning paradigm, we will also develop a generalized updating rule for policy gradient search in order to constantly improve learning performance. Our new updating rule has been examined on several benchmark learning problems. The experimental results on two problems will be reported specifically in this paper. Our results show that, under suitable generalization of the updating rule, the learning performance and reliability can be noticeably improved.

Conventional tracking methods are incapable of tracking boats towed by vehicles on boat ramps because the relative geometry of these combined objects changes as they move up and down the ramp. In the context of maritime boat ramp surveillance, fishing trailer boat is the object of interest for monitoring the amount of recreational fishing activities over the time. Instead of tracking trailer boat as a single object, this paper proposes a novel boat-vehicle combo object model, by which each boat is tracked as a combination of a trailered boat and a towing vehicle, and the relationship between these two components is modelled in multi-feature space and traced across consecutive frames. Experimental results show that the proposed combo modelling tracks the object of interest accurately and reliably in real-world boat traffic videos.