Learning to perceive the world as articulated: an approach for hierarchical learning in sensory-motor systems

This paper describes how agents can learn an internal model of the world structurally by focusing on the problem of behavior-based articulation. We develop an on-line learning scheme—the so-called mixture of recurrent neural net (RNN) experts—in which a set of RNN modules become self-organized as experts on multiple levels, in order to account for the different categories of sensory-motor flow which the robot experiences. Autonomous switching of activated modules in the lower level actually represents the articulation of the sensory-motor flow. In the meantime, a set of RNNs in the higher level competes to learn the sequences of module switching in the lower level, by which articulation at a further, more abstract level can be achieved. The proposed scheme was examined through simulation experiments involving the navigation learning problem. Our dynamical system analysis clarified the mechanism of the articulation. The possible correspondence between the articulation mechanism and the attention switching mechanism in thalamo-cortical loops is also discussed.     

Towards representation of a perceptual color manifold using associative memory for color constancy

In this paper, we propose the concept of a manifold of color perception through empirical observation that the center-surround properties of images in a perceptually similar environment define a manifold in the high dimensional space. Such a manifold representation can be learned using a novel recurrent neural network based learning algorithm. Unlike the conventional recurrent neural network model in which the memory is stored in an attractive fixed point at discrete locations in the state space, the dynamics of the proposed learning algorithm represent memory as a nonlinear line of attraction. The region of convergence around the nonlinear line is defined by the statistical characteristics of the training data. This learned manifold can then be used as a basis for color correction of the images having different color perception to the learned color perception. Experimental results show that the proposed recurrent neural network learning algorithm is capable of color balance the lighting variations in images captured in different environments successfully.     

Bayesian model search for mixture models based on optimizing variational bounds

When learning a mixture model, we suffer from the local optima and model structure determination problems. In this paper, we present a method for simultaneously solving these problems based on the variational Bayesian (VB) framework. First, in the VB framework, we derive an objective function that can simultaneously optimize both model parameter distributions and model structure. Next, focusing on mixture models, we present a deterministic algorithm to approximately optimize the objective function by using the idea of the split and merge operations which we previously proposed within the maximum likelihood framework. Then, we apply the method to mixture of expers (MoE) models to experimentally show that the proposed method can find the optimal number of experts of a MoE while avoiding local maxima.     

Neocognitron capable of incremental learning.

This paper proposes a new neocognitron that accepts incremental learning, without giving a severe damage to old memories or reducing learning speed. The new neocognitron uses a competitive learning, and the learning of all stages of the hierarchical network progresses simultaneously. To increase the learning speed, conventional neocognitrons of recent versions sacrificed the ability of incremental learning, and used a technique of sequential construction of layers, by which the learning of a layer started after the learning of the preceding layers had completely finished. If the learning speed is simply set high for the conventional neocognitron, simultaneous construction of layers produces many garbage cells, which become always silent after having finished the learning. The proposed neocognitron with a new learning method can prevent the generation of such garbage cells even with a high learning speed, allowing incremental learning.     

Sensitivity analysis applied to the construction of radial basis function networks.

Conventionally, a radial basis function (RBF) network is constructed by obtaining cluster centers of basis function by maximum likelihood learning. This paper proposes a novel learning algorithm for the construction of radial basis function using sensitivity analysis. In training, the number of hidden neurons and the centers of their radial basis functions are determined by the maximization of the output's sensitivity to the training data. In classification, the minimal number of such hidden neurons with the maximal sensitivity will be the most generalizable to unknown data. Our experimental results show that our proposed sensitivity-based RBF classifier outperforms the conventional RBFs and is as accurate as support vector machine (SVM). Hence, sensitivity analysis is expected to be a new alternative way to the construction of RBF networks.     

Extraction of primitive representation from captured human movements and measured ground reaction force to generate physically consistent imitated behaviors

In this paper, we propose an imitation learning framework to generate physically consistent behaviors by estimating the ground reaction force from captured human behaviors. In the proposed framework, we first extract behavioral primitives, which are represented by linear dynamical models, from captured human movements and measured ground reaction force by using the Gaussian mixture of linear dynamical models. Therefore, our method has small dependence on classification criteria defined by an experimenter. By switching primitives with different combinations while estimating the ground reaction force, different physically consistent behaviors can be generated. We apply the proposed method to a four-link robot model to generate squat motion sequences. The four-link robot model successfully generated the squat movements by using our imitation learning framework. To show generalization performance, we also apply the proposed method to robot models that have different torso weights and lengths from a human demonstrator and evaluate the control performances. In addition, we show that the robot model is able to recognize and imitate demonstrator movements even when the observed movements are deviated from the movements that are used to construct the primitives. For further evaluation in higher-dimensional state space, we apply the proposed method to a seven-link robot model. The seven-link robot model was able to generate squat-and-sway motions by using the proposed framework.     

An intelligent sales forecasting system through integration of artificial neural networks and fuzzy neural networks with fuzzy weight elimination.

Sales forecasting plays a very prominent role in business strategy. Numerous investigations addressing this problem have generally employed statistical methods, such as regression or autoregressive and moving average (ARMA). However, sales forecasting is very complicated owing to influence by internal and external environments. Recently, artificial neural networks (ANNs) have also been applied in sales forecasting since their promising performances in the areas of control and pattern recognition. However, further improvement is still necessary since unique circumstances, e.g. promotion, cause a sudden change in the sales pattern. Thus, this study utilizes a proposed fuzzy neural network (FNN), which is able to eliminate the unimportant weights, for the sake of learning fuzzy IF-THEN rules obtained from the marketing experts with respect to promotion. The result from FNN is further integrated with the time series data through an ANN. Both the simulated and real-world problem results show that FNN with weight elimination can have lower training error compared with the regular FNN. Besides, real-world problem results also indicate that the proposed estimation system outperforms the conventional statistical method and single ANN in accuracy.     

U2F-GAN: Weakly Supervised Super-pixel Segmentation in Thyroid Ultrasound Images

Precise nodule segmentation in thyroid ultrasound images is important for clinical quantitative analysis and diagnosis. Fully supervised deep learning method can effectively extract representative features from nodules and background. Despite the great success, deep learning–based segmentation methods still face a critical hindrance: the difficulty in acquiring sufficient training data due to high annotation costs. To this end, we propose a weakly supervised framework called uncertainty to fine generative adversarial network (U2F-GAN) for nodule segmentation in thyroid ultrasound images that exploits only a handful of rough bounding box annotations to successfully generate reliable labels from these weak supervisions. Based on feature-matching GAN, the proposed method alternates between generating masks and learning a segmentation network in an adversarial manner. Super-pixel processing mechanism is adopted to reflect low-level image structure features for learning and inferring semantic segmentation, which largely improve the efficiency of training process. In addition, we introduce a similarity comparison module and a distributed loss function with constraints to effectively remove noise in localization annotations and enhance the generalization capability of the network, thus strengthen the overall segmentation performance. Compared to existing weakly supervised approaches, our proposed U2F-GAN yields a significant performance boost. The segmentation results are also comparable to fully supervised methods, but the annotation burden is much lower. Also, the training speed of the network model is much faster than other methods with weak supervisions, which enables the network to be updated in time, thus is beneficial to high-throughput medical image setting.

     

Learning to imitate stochastic time series in a compositional way by chaos

This study shows that a mixture of RNN experts model can acquire the ability to generate sequences that are combination of multiple primitive patterns by means of self-organizing chaos. By training the model, each expert learns a primitive sequence pattern, and a gating network learns to imitate stochastic switching of the multiple primitives via chaotic dynamics, utilizing a sensitive dependence on initial conditions. As a demonstration, we present a numerical simulation in which the model learns Markov chain switching among some Lissajous curves by chaotic dynamics. Our analysis shows that by using a sufficient amount of training data, balanced with the network memory capacity, it is possible to satisfy the conditions for embedding the target stochastic sequences into a chaotic dynamical system. It is also shown that reconstruction of a stochastic time series by a chaotic model can be stabilized by adding a negligible amount of noise to the dynamics of the model.     

A Multi-Granularity Information-Based Method for Learning High-Dimensional Bayesian Network Structures

The purpose of structure learning is to construct a qualitative relationship of Bayesian networks. Bayesian network with interpretability and logicality is widely applied in a lot of fields. With the extensive development of high-dimensional and low sample size data in some applications, structure learning of Bayesian networks for high dimension and low sample size data becomes a challenging problem. To handle this problem, we propose a method for learning high-dimensional Bayesian network structures based on multi-granularity information. First, an undirected independence graph construction method containing global structure information is designed to optimize the search space of network structure. Then, an improved agglomerative hierarchical clustering method is presented to cluster variables into sub-granules, which reduces the complexity of structure learning by considering the variable community characteristic in high-dimensional data. Finally, the corresponding sub-graphs are formed by learning the internal structure of sub-granules, and the final network structure is constructed based on the proposed construct link graph algorithm. To verify the proposed method, we conduct two types of comparison experiments: comparison experiment and embedded comparison experiment. The results of the experiments show that our approach is superior to the competitors. The results indicate that our method can not only learn structures of Bayesian network from high-dimensional data efficiently but also improve the efficiency and accuracy of network structure generated by other algorithms for high-dimensional data.

     

Domain general and domain preferential brain regions associated with different types of task switching: a meta-analysis

One of our highest evolved functions as human beings is our capacity to switch between multiple tasks effectively. A body of research has identified a distributed frontoparietal network of brain regions which contribute to task switching. However, relatively less is known about whether some brain regions may contribute to switching in a domain-general manner while others may be more preferential for different kinds of switching. To explore this issue, we conducted three meta-analyses focusing on different types of task switching frequently used in the literature (perceptual, response, and context switching), and created a conjunction map of these distinct switch types. A total of 36 switching studies with 562 activation coordinates were analyzed using the activation likelihood estimation method. Common areas associated with switching across switch type included the inferior frontal junction and posterior parietal cortex. In contrast, domain-preferential activation was observed for perceptual switching in the dorsal portion of the premotor cortex and for context switching in frontopolar cortex. Our results suggest that some regions within the frontoparietal network contribute to domain-general switching processes while others contribute to more domain-preferential processes, according to the type of task switch performed.     

Event-triggered synchronization strategy for complex dynamical networks with the Markovian switching topologies

This paper concerns the synchronization problem of complex networks with the random switching topologies. By modeling the switching of network topologies as a Markov process, a novel event-triggered synchronization strategy is proposed. Unlike the existing strategies, the event detection of this strategy only works at the network topology switching time instant, which can significantly decrease the communication frequency between nodes and save the network resources. Under this strategy, the synchronization problem of complex network is equivalently converted to the stability of a class of Markovian jump systems with a time-varying delay. By using the Lyapunov–Krasovskii functional method and the weak infinitesimal operation, a sufficient condition for the mean square synchronization of the complex networks subject to Markovian switching topologies is established. Finally, a numerical simulation example is provided to demonstrate the theoretical results.     

Characterising switching behaviour in perceptual multi-stability

When people experience an unchanging sensory input for a long period of time, their perception tends to switch stochastically and unavoidably between alternative interpretations of the sensation; a phenomenon known as perceptual bi-stability or multi-stability. The huge variability in the experimental data obtained in such paradigms makes it difficult to distinguish typical patterns of behaviour, or to identify differences between switching patterns. Here we propose a new approach to characterising switching behaviour based upon the extraction of transition matrices from the data, which provide a compact representation that is well-understood mathematically. On the basis of this representation we can characterise patterns of perceptual switching, visualise and simulate typical switching patterns, and calculate the likelihood of observing a particular switching pattern. The proposed method can support comparisons between different observers, experimental conditions and even experiments. We demonstrate the insights offered by this approach using examples from our experiments investigating multi-stability in auditory streaming. However, the methodology is generic and thus widely applicable in studies of multi-stability in any domain.     

Logarithmic learning for generalized classifier neural network

Generalized classifier neural network is introduced as an efficient classifier among the others. Unless the initial smoothing parameter value is close to the optimal one, generalized classifier neural network suffers from convergence problem and requires quite a long time to converge. In this work, to overcome this problem, a logarithmic learning approach is proposed. The proposed method uses logarithmic cost function instead of squared error. Minimization of this cost function reduces the number of iterations used for reaching the minima. The proposed method is tested on 15 different data sets and performance of logarithmic learning generalized classifier neural network is compared with that of standard one. Thanks to operation range of radial basis function included by generalized classifier neural network, proposed logarithmic approach and its derivative has continuous values. This makes it possible to adopt the advantage of logarithmic fast convergence by the proposed learning method. Due to fast convergence ability of logarithmic cost function, training time is maximally decreased to 99.2%. In addition to decrease in training time, classification performance may also be improved till 60%. According to the test results, while the proposed method provides a solution for time requirement problem of generalized classifier neural network, it may also improve the classification accuracy. The proposed method can be considered as an efficient way for reducing the time requirement problem of generalized classifier neural network.     

Sparse temporally dynamic resting-state functional connectivity networks for early MCI identification

In conventional resting-state functional MRI (R-fMRI) analysis, functional connectivity is assumed to be temporally stationary, overlooking neural activities or interactions that may happen within the scan duration. Dynamic changes of neural interactions can be reflected by variations of topology and correlation strength in temporally correlated functional connectivity networks. These connectivity networks may potentially capture subtle yet short neural connectivity disruptions induced by disease pathologies. Accordingly, we are motivated to utilize disrupted temporal network properties for improving control-patient classification performance. Specifically, a sliding window approach is firstly employed to generate a sequence of overlapping R-fMRI sub-series. Based on these sub-series, sliding window correlations, which characterize the neural interactions between brain regions, are then computed to construct a series of temporal networks. Individual estimation of these temporal networks using conventional network construction approaches fails to take into consideration intrinsic temporal smoothness among successive overlapping R-fMRI sub-series. To preserve temporal smoothness of R-fMRI sub-series, we suggest to jointly estimate the temporal networks by maximizing a penalized log likelihood using a fused sparse learning algorithm. This sparse learning algorithm encourages temporally correlated networks to have similar network topology and correlation strengths. We design a disease identification framework based on the estimated temporal networks, and group level network property differences and classification results demonstrate the importance of including temporally dynamic R-fMRI scan information to improve diagnosis accuracy of mild cognitive impairment patients.     

Learning the pseudoinverse solution to network weights

The last decade has seen the parallel emergence in computational neuroscience and machine learning of neural network structures which spread the input signal randomly to a higher dimensional space; perform a nonlinear activation; and then solve for a regression or classification output by means of a mathematical pseudoinverse operation. In the field of neuromorphic engineering, these methods are increasingly popular for synthesizing biologically plausible neural networks, but the “learning method”—computation of the pseudoinverse by singular value decomposition—is problematic both for biological plausibility and because it is not an online or an adaptive method. We present an online or incremental method of computing the pseudoinverse precisely, which we argue is biologically plausible as a learning method, and which can be made adaptable for non-stationary data streams. The method is significantly more memory-efficient than the conventional computation of pseudoinverses by singular value decomposition.     

A global learning algorithm for a RBF network

This article presents a new learning algorithm for the construction and training of a RBF neural network. The algorithm is based on a global mechanism of parameter learning using a maximum likelihood classification approach. The resulting neurons in the RBF network partitions a multidimensional pattern space into a set of maximum-size hyper-ellipsoid subspaces in terms of the statistical distributions of the training samples. An important feature of the algorithm is that the learning process includes both the tasks of discovering a suitable network structure and of determining the connection weights. The entire network and its parameters are thought of evolved gradually in the learning process.     

An improved time series prediction by applying the layer-by-layer learning method to FIR neural networks

The FIR neural network model was recently proposed for time series prediction and gave good results. However, the learning algorithm used for the FIR network is a kind of gradient descent method and hence inherits all the well-known problems of the method. Recently a new learning algorithm called the optimization layer by layer was proposed for the regular multilayer perceptron network, and showed a great improvement in the learning time as well as the performance of the network.

In this paper we develop a new learning algorithm for the FIR neural network model by applying the idea of the optimization layer by layer to the model. The results of the experiment, using two popular time series prediction problems, show that the new algorithm is far better in learning time and more accurate in prediction performance than the original learning algorithm.     

Acquisition of nonlinear forward optics in generative models: Two-stage “downside-up” learning for occluded vision

We propose a two-stage learning method which implements occluded visual scene analysis into a generative model, a type of hierarchical neural network with bi-directional synaptic connections. Here, top-down connections simulate forward optics to generate predictions for sensory driven low-level representation, whereas bottom-up connections function to send the prediction error, the difference between the sensory based and the predicted low-level representation, to higher areas. The prediction error is then used to update the high-level representation to obtain better agreement with the visual scene. Although the actual forward optics is highly nonlinear and the accuracy of simulated forward optics is crucial for these types of models, the majority of previous studies have only investigated linear and simplified cases of forward optics. Here we take occluded vision as an example of nonlinear forward optics, where an object in front completely masks out the object behind. We propose a two-staged learning method inspired by the staged development of infant visual capacity. In the primary learning stage, a minimal set of object basis is acquired within a linear generative model using the conventional unsupervised learning scheme. In the secondary learning stage, an auxiliary multi-layer neural network is trained to acquire nonlinear forward optics by supervised learning. The important point is that the high-level representation of the linear generative model serves as the input and the sensory driven low-level representation provides the desired output. Numerical simulations show that occluded visual scene analysis can indeed be implemented by the proposed method. Furthermore, considering the format of input to the multi-layer network and analysis of hidden-layer units leads to the prediction that whole object representation of partially occluded objects, together with complex intermediate representation as a consequence of nonlinear transformation from non-occluded to occluded representation may exist in the low-level visual system of the brain.     

Functional network estimation using multigraph learning with application to brain maturation study

Although most dramatic structural changes occur in the perinatal period, a growing body of evidences demonstrates that adolescence and early adulthood are also important for substantial neurodevelopment. We were thus motivated to explore brain development during puberty by evaluating functional connectivity network (FCN) differences between childhood and young adulthood using multi‐paradigm task‐based functional magnetic resonance imaging (fMRI) measurements. Different from conventional multigraph based FCN construction methods where the graph network was built independently for each modality/paradigm, we proposed a multigraph learning model in this work. It promises a better fitting to FCN construction by jointly estimating brain network from multi‐paradigm fMRI time series, which may share common graph structures. To investigate the hub regions of the brain, we further conducted graph Fourier transform (GFT) to divide the fMRI BOLD time series of a node within the brain network into a range of frequencies. Then we identified the hub regions characterizing brain maturity through eigen‐analysis of the low frequency components, which were believed to represent the organized structures shared by a large population. The proposed method was evaluated using both synthetic and real data, which demonstrated its effectiveness in extracting informative brain connectivity patterns. We detected 14 hub regions from the child group and 12 hub regions from the young adult group. We show the significance of these findings with a discussion of their functions and activation patterns as a function of age. In summary, our proposed method can extract brain connectivity network more accurately by considering the latent common structures between different fMRI paradigms, which are significant for both understanding brain development and recognizing population groups of different ages.