Implementing Gaussian process inference with neural networks

Gaussian processes compare favourably with backpropagation neural networks as a tool for regression, and Bayesian neural networks have Gaussian process behaviour when the number of hidden neurons tends to infinity. We describe a simple recurrent neural network with connection weights trained by one-shot Hebbian learning. This network amounts to a dynamical system which relaxes to a stable state in which it generates predictions identical to those of Gaussian process regression. In effect an infinite number of hidden units in a feed-forward architecture can be replaced by a merely finite number, together with recurrent connections.     

Self-selected modular recurrent neural networks with postural and inertial subnetworks applied to complex movements

 It has been shown that dynamic recurrent neural networks are successful in identifying the complex mapping relationship between full-wave-rectified electromyographic (EMG) signals and limb trajectories during complex movements. These connectionist models include two types of adaptive parameters: the interconnection weights between the units and the time constants associated to each neuron-like unit; they are governed by continuous-time equations. Due to their internal structure, these models are particularly appropriate to solve dynamical tasks (with time-varying input and output signals). We show in this paper that the introduction of a modular organization dedicated to different aspects of the dynamical mapping including privileged communication channels can refine the architecture of these recurrent networks. We first divide the initial individual network into two communicating subnetworks. These two modules receive the same EMG signals as input but are involved in different identification tasks related to position and acceleration. We then show that the introduction of an artificial distance in the model (using a Gaussian modulation factor of weights) induces a reduced modular architecture based on a self-elimination of null synaptic weights. Moreover, this self-selected reduced model based on two subnetworks performs the identification task better than the original single network while using fewer free parameters (better learning curve and better identification quality). We also show that this modular network exhibits several features that can be considered as biologically plausible after the learning process: self-selection of a specific inhibitory communicating path between both subnetworks after the learning process, appearance of tonic and phasic neurons, and coherent distribution of the values of the time constants within each subnetwork. Received: 17 September 2001 / Accepted in revised form: 15 January 2002     

Fine-tuning and the stability of recurrent neural networks

A central criticism of standard theoretical approaches to constructing stable, recurrent model networks is that the synaptic connection weights need to be finely-tuned. This criticism is severe because proposed rules for learning these weights have been shown to have various limitations to their biological plausibility. Hence it is unlikely that such rules are used to continuously fine-tune the network in vivo. We describe a learning rule that is able to tune synaptic weights in a biologically plausible manner. We demonstrate and test this rule in the context of the oculomotor integrator, showing that only known neural signals are needed to tune the weights. We demonstrate that the rule appropriately accounts for a wide variety of experimental results, and is robust under several kinds of perturbation. Furthermore, we show that the rule is able to achieve stability as good as or better than that provided by the linearly optimal weights often used in recurrent models of the integrator. Finally, we discuss how this rule can be generalized to tune a wide variety of recurrent attractor networks, such as those found in head direction and path integration systems, suggesting that it may be used to tune a wide variety of stable neural systems.     

Constructive Autoassociative Neural Network for Facial Recognition

Autoassociative artificial neural networks have been used in many different computer vision applications. However, it is difficult to define the most suitable neural network architecture because this definition is based on previous knowledge and depends on the problem domain. To address this problem, we propose a constructive autoassociative neural network called CANet (Constructive Autoassociative Neural Network). CANet integrates the concepts of receptive fields and autoassociative memory in a dynamic architecture that changes the configuration of the receptive fields by adding new neurons in the hidden layer, while a pruning algorithm removes neurons from the output layer. Neurons in the CANet output layer present lateral inhibitory connections that improve the recognition rate. Experiments in face recognition and facial expression recognition show that the CANet outperforms other methods presented in the literature.     

Functional identification of biological neural networks using reservoir adaptation for point processes

The complexity of biological neural networks does not allow to directly relate their biophysical properties to the dynamics of their electrical activity. We present a reservoir computing approach for functionally identifying a biological neural network, i.e. for building an artificial system that is functionally equivalent to the reference biological network. Employing feed-forward and recurrent networks with fading memory, i.e. reservoirs, we propose a point process based learning algorithm to train the internal parameters of the reservoir and the connectivity between the reservoir and the memoryless readout neurons. Specifically, the model is an Echo State Network (ESN) with leaky integrator neurons, whose individual leakage time constants are also adapted. The proposed ESN algorithm learns a predictive model of stimulus-response relations in in vitro and simulated networks, i.e. it models their response dynamics. Receiver Operating Characteristic (ROC) curve analysis indicates that these ESNs can imitate the response signal of a reference biological network. Reservoir adaptation improved the performance of an ESN over readout-only training methods in many cases. This also held for adaptive feed-forward reservoirs, which had no recurrent dynamics. We demonstrate the predictive power of these ESNs on various tasks with cultured and simulated biological neural networks.     

Dense Neuron Clustering Explains Connectivity Statistics in Cortical Microcircuits

Local cortical circuits appear highly non-random, but the underlying connectivity rule remains elusive. Here, we analyze experimental data observed in layer 5 of rat neocortex and suggest a model for connectivity from which emerge essential observed non-random features of both wiring and weighting. These features include lognormal distributions of synaptic connection strength, anatomical clustering, and strong correlations between clustering and connection strength. Our model predicts that cortical microcircuits contain large groups of densely connected neurons which we call clusters. We show that such a cluster contains about one fifth of all excitatory neurons of a circuit which are very densely connected with stronger than average synapses. We demonstrate that such clustering plays an important role in the network dynamics, namely, it creates bistable neural spiking in small cortical circuits. Furthermore, introducing local clustering in large-scale networks leads to the emergence of various patterns of persistent local activity in an ongoing network activity. Thus, our results may bridge a gap between anatomical structure and persistent activity observed during working memory and other cognitive processes.     

Improving the prediction accuracy of residue solvent accessibility and real-value backbone torsion angles of proteins by guided-learning through a two-layer neural network

This article attempts to increase the prediction accuracy of residue solvent accessibility and real-value backbone torsion angles of proteins through improved learning. Most methods developed for improving the backpropagation algorithm of artificial neural networks are limited to small neural networks. Here, we introduce a guided-learning method suitable for networks of any size. The method employs a part of the weights for guiding and the other part for training and optimization. We demonstrate this technique by predicting residue solvent accessibility and real-value backbone torsion angles of proteins. In this application, the guiding factor is designed to satisfy the intuitive condition that for most residues, the contribution of a residue to the structural properties of another residue is smaller for greater separation in the protein-sequence distance between the two residues. We show that the guided-learning method makes a 2-4% reduction in 10-fold cross-validated mean absolute errors (MAE) for predicting residue solvent accessibility and backbone torsion angles, regardless of the size of database, the number of hidden layers and the size of input windows. This together with introduction of two-layer neural network with a bipolar activation function leads to a new method that has a MAE of 0.11 for residue solvent accessibility, 36 degrees for psi, and 22 degrees for phi. The method is available as a Real-SPINE 3.0 server in http://sparks.informatics.iupui.edu.     

人工神经网络模型及其在农业和生态学研究中的应用

对于一些复杂的农业生态系统,人们对其生态过程了解较少,且这些系统的不确定性和模糊性较大,用传统的方法难以模拟这些系统的行为,神经网络模型因为能较精确地模拟这些系统的行为,而引起生态学者们的广泛兴趣。该文着重介绍了误差逆传神经网络模型的结构、算法及其在农业和生态学中的应用研究。误差逆传神经网络模型一般采用三层神经网络模型结构,三层的神经网络模型能模拟任意复杂程度的连续函数,而且因为它的结构小而不容易产生与训练数据的过度吻合。误差逆传神经网络模型算法的主要特征是:利用当前的输入误差对权值进行调整。在生态学和农业研究中,误差逆传神经网络模型通常作为非线性函数模拟器用于预测作物产量、生物生产量、生物与环境之间的关系等。已有的研究表明:误差逆传神经网络模型的模拟精度要远远高于多元线性方程,类似于非线性方程,而在样本量足够的情况下,有一定的外推能力。但是误差逆传神经网络模型需要大量的样本量来保证所求取参数的可靠性,但这在实际研究中很难做到,因而限制了误差逆传神经网络模型的应用。近年来人们提出了强制训练停止、复合模型等多种技术来提高误差逆传神经网络模型的外推能力,也提出了Garson算法、敏感性分析以及随机化检验等技术对误差逆传神经网络模型的机理进行解释。误差逆传神经网络模型的真正优势在于模拟人们了解较少或不确定性和模糊性较大系统的行为,这些是传统模型所无法实现的,因而是对传统机理模型的重要补充。     

Some structural determinants of Pavlovian conditioning in artificial neural networks

This paper investigates the possible role of neuroanatomical features in Pavlovian conditioning, via computer simulations with layered, feedforward artificial neural networks. The networks’ structure and functioning are described by a strongly bottom-up model that takes into account the roles of hippocampal and dopaminergic systems in conditioning. Neuroanatomical features were simulated as generic structural or architectural features of neural networks. We focused on the number of units per hidden layer and connectivity. The effect of the number of units per hidden layer was investigated through simulations of resistance to extinction in fully connected networks. Large networks were more resistant to extinction than small networks, a stochastic effect of the asynchronous random procedure used in the simulator to update activations and weights. These networks did not simulate second-order conditioning because weight competition prevented conditioning to a stimulus after conditioning to another. Partially connected networks simulated second-order conditioning and devaluation of the second-order stimulus after extinction of a similar first-order stimulus. Similar stimuli were simulated as nonorthogonal input-vectors.     

Adaptive Neural Network Structure Optimization Algorithm Based on Dynamic Nodes

Large-scale artificial neural networks have many redundant structures, making the network fall into the issue of local optimization and extended training time. Moreover, existing neural network topology optimization algorithms have the disadvantage of many calculations and complex network structure modeling. We propose a Dynamic Node-based neural network Structure optimization algorithm (DNS) to handle these issues. DNS consists of two steps: the generation step and the pruning step. In the generation step, the network generates hidden layers layer by layer until accuracy reaches the threshold. Then, the network uses a pruning algorithm based on Hebb’s rule or Pearson’s correlation for adaptation in the pruning step. In addition, we combine genetic algorithm to optimize DNS (GA-DNS). Experimental results show that compared with traditional neural network topology optimization algorithms, GA-DNS can generate neural networks with higher construction efficiency, lower structure complexity, and higher classification accuracy.     

A biologically inspired neural network for dynamic programming

An artificial neural network with a two-layer feedback topology and generalized recurrent neurons, for solving nonlinear discrete dynamic optimization problems, is developed. A direct method to assign the weights of neural networks is presented. The method is based on Bellmann's Optimality Principle and on the interchange of information which occurs during the synaptic chemical processing among neurons. The neural network based algorithm is an advantageous approach for dynamic programming due to the inherent parallelism of the neural networks; further it reduces the severity of computational problems that can occur in methods like conventional methods. Some illustrative application examples are presented to show how this approach works out including the shortest path and fuzzy decision making problems.     

Associative memory in quaternionic Hopfield neural network

Associative memory networks based on quaternionic Hopfield neural network are investigated in this paper. These networks are composed of quaternionic neurons, and input, output, threshold, and connection weights are represented in quaternions, which is a class of hypercomplex number systems. The energy function of the network and the Hebbian rule for embedding patterns are introduced. The stable states and their basins are explored for the networks with three neurons and four neurons. It is clarified that there exist at most 16 stable states, called multiplet components, as the degenerated stored patterns, and each of these states has its basin in the quaternionic networks.     

An extended model for a spiking neuron class

This paper proposes an extension to the model of a spiking neuron for information processing in artificial neural networks, developing a new approach for the dynamic threshold of the integrate-and-fire neuron. This new approach invokes characteristics of biological neurons such as the behavior of chemical synapses and the receptor field. We demonstrate how such a digital model of spiking neurons can solve complex nonlinear classification with a single neuron, performing experiments for the classical XOR problem. Compared with rate-coded networks and the classical integrate-and-fire model, the trained network demonstrated faster information processing, requiring fewer neurons and shorter learning periods. The extended model validates all the logic functions of biological neurons when such functions are necessary for the proper flow of binary codes through a neural network.     

Resolution of binary mixtures of microorganisms using electrochemical impedance spectroscopy and artificial neural networks

This work describes the resolution of binary mixtures of microorganisms using electrochemical impedance spectroscopy (EIS) and artificial neural networks (ANNs) for the processing of data. Pseudomonas aeruginosa, Staphylococcus aureus and Saccharomyces cerevisiae were chosen as models for Gram-negative bacteria, Gram-positive bacteria and yeasts, respectively. In this study, best results were obtained when entering the imaginary component of the impedance at each frequency (strongly related to the capacitive elements of the electrical equivalent circuit) into backpropagation neural networks made up by two hidden layers. The optimal configuration of these layers respectively used the radbas and the logsig transfer functions with 4 or 6 neurons in the first hidden layer and 10 neurons in the second one. In all cases, good prediction ability was obtained with correlation coefficients better than 0.989 when comparing the predicted and the expected values for a set of six external test samples not used in the training process.     

Short-term prediction of chaotic time series by using RBF network with regression weights

We propose a framework for constructing and training a radial basis function (RBF) neural network. The structure of the gaussian functions is modified using a pseudo-gaussian function (PG) in which two scaling parameters sigma are introduced, which eliminates the symmetry restriction and provides the neurons in the hidden layer with greater flexibility with respect to function approximation. We propose a modified PG-BF (pseudo-gaussian basis function) network in which the regression weights are used to replace the constant weights in the output layer. For this purpose, a sequential learning algorithm is presented to adapt the structure of the network, in which it is possible to create a new hidden unit and also to detect and remove inactive units. A salient feature of the network systems is that the method used for calculating the overall output is the weighted average of the output associated with each receptive field. The superior performance of the proposed PG-BF system over the standard RBF are illustrated using the problem of short-term prediction of chaotic time series.     

Bayesian Computation Emerges in Generic Cortical Microcircuits through Spike-Timing-Dependent Plasticity

The principles by which networks of neurons compute, and how spike-timing dependent plasticity (STDP) of synaptic weights generates and maintains their computational function, are unknown. Preceding work has shown that soft winner-take-all (WTA) circuits, where pyramidal neurons inhibit each other via interneurons, are a common motif of cortical microcircuits. We show through theoretical analysis and computer simulations that Bayesian computation is induced in these network motifs through STDP in combination with activity-dependent changes in the excitability of neurons. The fundamental components of this emergent Bayesian computation are priors that result from adaptation of neuronal excitability and implicit generative models for hidden causes that are created in the synaptic weights through STDP. In fact, a surprising result is that STDP is able to approximate a powerful principle for fitting such implicit generative models to high-dimensional spike inputs: Expectation Maximization. Our results suggest that the experimentally observed spontaneous activity and trial-to-trial variability of cortical neurons are essential features of their information processing capability, since their functional role is to represent probability distributions rather than static neural codes. Furthermore it suggests networks of Bayesian computation modules as a new model for distributed information processing in the cortex.     

Information maintenance and statistical dependence reduction in simple neural networks

This study compares the ability of excitatory, feed-forward neural networks to construct good transformations on their inputs. The quality of such a transformation is judged by the minimization of two information measures: the information loss of the transformation and the statistical dependency of the output. The networks that are compared differ from each other in the parametric properties of their neurons and in their connectivity. The particular network parameters studied are output firing threshold, synaptic connectivity, and associative modification of connection weights. The network parameters that most directly affect firing levels are threshold and connectivity. Networks incorporating neurons with dynamic threshold adjustment produce better transformations. When firing threshold is optimized, sparser synaptic connectivity produces a better transformation than denser connectivity. Associative modification of synaptic weights confers only a slight advantage in the construction of optimal transformations. Additionally, our research shows that some environments are better suited than others for recoding. Specifically, input environments high in statistical dependence, i.e. those environments most in need of recoding, are more likely to undergo successful transformations.     

Liquid state machines and cultured cortical networks: the separation property

In vitro neural networks of cortical neurons interfaced to a computer via multichannel microelectrode arrays (MEA) provide a unique paradigm to create a hybrid neural computer. Unfortunately, only rudimentary information about these in vitro network's computational properties or the extent of their abilities are known. To study those properties, a liquid state machine (LSM) approach was employed in which the liquid (typically an artificial neural network) was replaced with a living cortical network and the input and readout functions were replaced by the MEA-computer interface. A key requirement of the LSM architecture is that inputs into the liquid state must result in separable outputs based on the liquid's response (separation property). In this paper, high and low frequency multi-site stimulation patterns were applied to the living cortical networks. Two template-based classifiers, one based on Euclidean distance and a second based on a cross-correlation were then applied to measure the separation of the input-output relationship. The result was over a 95% (99.8% when nonstationarity is compensated) input reconstruction accuracy for the high and low frequency patterns, confirming the existence of the separation property in these biological networks.