A local training-pruning approach for recurrent neural networks

The global extended Kalman filtering (EKF) algorithm for recurrent neural networks (RNNs) is plagued by the drawback of high computational cost and storage requirement. In this paper, we present a local EKF training-pruning approach that can solve this problem. In particular, the by-products, obtained along with the local EKF training, can be utilized to measure the importance of the network weights. Comparing with the original global approach, the proposed local approach results in much lower computational cost and storage requirement. Hence, it is more practical in solving real world problems. Simulation showed that our approach is an effective joint-training-pruning method for RNNs under online operation.     

Universal distribution of saliencies for pruning in layered neural networks

A better understanding of pruning methods based on a ranking of weights according to their saliency in a trained network requires further information on the statistical properties of such saliencies. We focus on two-layer networks with either a linear or nonlinear output unit, and obtain analytic expressions for the distribution of saliencies and their logarithms. Our results reveal unexpected universal properties of the log-saliency distribution and suggest a novel algorithm for saliency-based weight ranking that avoids the numerical cost of second derivative evaluations.     

A cascading structure and training method for multilayer neural networks

A new structure and training method for multilayer neural networks is presented. The proposed method is based on cascade training of subnetworks and optimizing weights layer by layer. The training procedure is completed in two steps. First, a subnetwork, m inputs and n outputs as the style of training samples, is trained using the training samples. Secondly the outputs of the subnetwork is taken as the inputs and the outputs of the training sample as the desired outputs, another subnetwork with n inputs and n outputs is trained. Finally the two trained subnetworks are connected and a trained multilayer neural networks is created. The numerical simulation results based on both linear least squares back-propagation (LSB) and traditional back-propagation (BP) algorithm have demonstrated the efficiency of the proposed method.     

Dynamics of pruning in simulated large-scale spiking neural networks

Massive synaptic pruning following over-growth is a general feature of mammalian brain maturation. This article studies the synaptic pruning that occurs in large networks of simulated spiking neurons in the absence of specific input patterns of activity. The evolution of connections between neurons were governed by an original bioinspired spike-timing-dependent synaptic plasticity (STDP) modification rule which included a slow decay term. The network reached a steady state with a bimodal distribution of the synaptic weights that were either incremented to the maximum value or decremented to the lowest value. After 1x10(6) time steps the final number of synapses that remained active was below 10% of the number of initially active synapses independently of network size. The synaptic modification rule did not introduce spurious biases in the geometrical distribution of the remaining active projections. The results show that, under certain conditions, the model is capable of generating spontaneously emergent cell assemblies.     

A genetic algorithm with adaptive mutations and family competition for training neural networks

In this paper, we present a new evolutionary technique to train three general neural networks. Based on family competition principles and adaptive rules, the proposed approach integrates decreasing-based mutations and self-adaptive mutations to collaborate with each other. Different mutations act as global and local strategies respectively to balance the trade-off between solution quality and convergence speed. Our algorithm is then applied to three different task domains: Boolean functions, regular language recognition, and artificial ant problems. Experimental results indicate that the proposed algorithm is very competitive with comparable evolutionary algorithms. We also discuss the search power of our proposed approach.     

A Lempel-Ziv complexity-based neural network pruning algorithm

This paper presents a pruning method for artificial neural networks (ANNs) based on the 'Lempel-Ziv complexity' (LZC) measure. We call this method the 'silent pruning algorithm' (SPA). The term 'silent' is used in the sense that SPA prunes ANNs without causing much disturbance during the network training. SPA prunes hidden units during the training process according to their ranks computed from LZC. LZC extracts the number of unique patterns in a time sequence obtained from the output of a hidden unit and a smaller value of LZC indicates higher redundancy of a hidden unit. SPA has a great resemblance to biological brains since it encourages higher complexity during the training process. SPA is similar to, yet different from, existing pruning algorithms. The algorithm has been tested on a number of challenging benchmark problems in machine learning, including cancer, diabetes, heart, card, iris, glass, thyroid, and hepatitis problems. We compared SPA with other pruning algorithms and we found that SPA is better than the 'random deletion algorithm' (RDA) which prunes hidden units randomly. Our experimental results show that SPA can simplify ANNs with good generalization ability.     

Sequential configuration model for firing patterns in local neural networks

This paper presents a sequential configuration model to represent the coordinated firing patterns of memory traces in groups of neurons in local networks. Computer simulations are used to study the dynamic properties of memory traces selectively retrieved from networks in which multiple memory traces have been embedded according to the sequential configuration model. Distinct memory traces which utilize the same neurons, but differ only in temporal sequencing are selectively retrievable. Firing patterns of constituent neurons of retrieved memory traces exhibit the main properties of neurons observed in multi microelectrode recordings. The paper shows how to adjust relative synaptic weightings so as to control the disruptive influences of cross-talk in multipy-embedded networks. The theoretical distinction between (primarily anatomical) beds and (primarily physiological) realizations underlines the fundamentally stochastic nature of network firing patterns, and allows the definition of 4 degrees of clarity of retrieved memory traces.     

A robust neural networks approach for spatial and intensity-dependent normalization of cDNA microarray data

MOTIVATION: Microarray experiments are affected by numerous sources of non-biological variation that contribute systematic bias to the resulting data. In a dual-label (two-color) cDNA or long-oligonucleotide microarray, these systematic biases are often manifested as an imbalance of measured fluorescent intensities corresponding to Sample A versus those corresponding to Sample B. Systematic biases also affect between-slide comparisons. Making effective corrections for these systematic biases is a requisite for detecting the underlying biological variation between samples. Effective data normalization is therefore an essential step in the confident identification of biologically relevant differences in gene expression profiles. Several normalization methods for the correction of systemic bias have been described. While many of these methods have addressed intensity-dependent bias, few have addressed both intensity-dependent and spatiality-dependent bias. RESULTS: We present a neural network-based normalization method for correcting the intensity- and spatiality-dependent bias in cDNA microarray datasets. In this normalization method, the dependence of the log-intensity ratio (M) on the average log-intensity (A) as well as on the spatial coordinates (X,Y) of spots is approximated with a feed-forward neural network function. Resistance to outliers is provided by assigning weights to each spot based on how distant their M values is from the median over the spots whose A values are similar, as well as by using pseudospatial coordinates instead of spot row and column indices. A comparison of the robust neural network method with other published methods demonstrates its potential in reducing both intensity-dependent bias and spatial-dependent bias, which translates to more reliable identification of truly regulated genes.     

A neural network approach to approximating MAP in belief networks

Bayesian belief networks (BBN) are a widely studied graphical model for representing uncertainty and probabilistic interdependence among variables. One of the factors that restricts the model's wide acceptance in practical applications is that the general inference with BBN is NP-hard. This is also true for the maximum a posteriori probability (MAP) problem, which is to find the most probable joint value assignment to all uninstantiated variables, given instantiation of some variables in a BBN. To circumvent the difficulty caused by MAP's computational complexity, we suggest in this paper a neural network approximation approach. With this approach, a BBN is treated as a neural network without any change or transformation of the network structure, and the node activation functions are derived based on an energy function defined over a given BBN. Three methods are developed. They are the hill-climbing style discrete method, the simulated annealing method, and the continuous method based on the mean field theory. All three methods are for BBN of general structures, with the restriction that nodes of BBN are binary variables. In addition, rules for applying these methods to noisy-or networks are also developed, which may lead to more efficient computation in some cases. These methods' convergence is analyzed, and their validity tested through a series of computer experiments with two BBN of moderate size and complexity. Although additional theoretical and empirical work is needed, the analysis and experiments suggest that this approach may lead to effective and accurate approximation for MAP problems.     

Entropy learning and relevance criteria for neural network pruning

In this paper, entropy is a term used in the learning phase of a neural network. As learning progresses, more hidden nodes get into saturation. The early creation of such hidden nodes may impair generalisation. Hence an entropy approach is proposed to dampen the early creation of such nodes by using a new computation called entropy cycle. Entropy learning also helps to increase the importance of relevant nodes while dampening the less important nodes. At the end of learning, the less important nodes can then be pruned to reduce the memory requirements of the neural network.     

Recent advances in the MOBJ algorithm for training artificial neural networks

This paper presents a new scheme for training MLPs which employs a relaxation method for multi-objective optimization. The algorithm works by obtaining a reduced set of solutions, from which the one with the best generalization is selected. This approach allows balancing between the training error and norm of network weight vectors, which are the two objective functions of the multi-objective optimization problem. The method is applied to classification and regression problems and compared with Weight Decay (WD), Support Vector Machines (SVMs) and standard Backpropagation (BP). It is shown that the systematic procedure for training proposed results on good generalization neural models, and outperforms traditional methods.     

Deterministic convergence of chaos injection-based gradient method for training feedforward neural networks

It has been shown that, by adding a chaotic sequence to the weight update during the training of neural networks, the chaos injection-based gradient method (CIBGM) is superior to the standard backpropagation algorithm. This paper presents the theoretical convergence analysis of CIBGM for training feedforward neural networks. We consider both the case of batch learning as well as the case of online learning. Under mild conditions, we prove the weak convergence, i.e., the training error tends to a constant and the gradient of the error function tends to zero. Moreover, the strong convergence of CIBGM is also obtained with the help of an extra condition. The theoretical results are substantiated by a simulation example.     

Effect of stimulus-driven pruning on the detection of spatiotemporal patterns of activity in large neural networks

Adult patterns of neuronal connectivity develop from a transient embryonic template characterized by exuberant projections to both appropriate and inappropriate target regions in a process known as synaptic pruning. Trigger signals able to induce synaptic pruning could be related to dynamic functions that depend on the timing of action potentials. We stimulated locally connected random networks of spiking neurons and observed the effect of a spike-timing-dependent synaptic plasticity (STDP)-driven pruning process on the emergence of cell assemblies. The spike trains of the simulated excitatory neurons were recorded. We searched for spatiotemporal firing patterns as potential markers of the build-up of functionally organized recurrent activity associated with spatially organized connectivity.     

A new training algorithm using artificial neural networks to classify gender-specific dynamic gait patterns

The aim of this study was to present a new training algorithm using artificial neural networks called multi-objective least absolute shrinkage and selection operator (MOBJ-LASSO) applied to the classification of dynamic gait patterns. The movement pattern is identified by 20 characteristics from the three components of the ground reaction force which are used as input information for the neural networks in gender-specific gait classification. The classification performance between MOBJ-LASSO (97.4%) and multi-objective algorithm (MOBJ) (97.1%) is similar, but the MOBJ-LASSO algorithm achieved more improved results than the MOBJ because it is able to eliminate the inputs and automatically select the parameters of the neural network. Thus, it is an effective tool for data mining using neural networks. From 20 inputs used for training, MOBJ-LASSO selected the first and second peaks of the vertical force and the force peak in the antero-posterior direction as the variables that classify the gait patterns of the different genders.     

Training radial basis function neural networks: effects of training set size and imbalanced training sets

Obtaining training data for constructing artificial neural networks (ANNs) to identify microbiological taxa is not always easy. Often, only small data sets with different numbers of observations per taxon are available. Here, the effect of both size of the training data set and of an imbalanced number of training patterns for different taxa is investigated using radial basis function ANNs to identify up to 60 species of marine microalgae. The best networks trained to discriminate 20, 40 and 60 species respectively gave overall percentage correct identification of 92, 84 and 77%. From 100 to 200 patterns per species was sufficient in networks trained to discriminate 20, 40 or 60 species. For 40 and 60 species data sets an imbalance in the number of training patterns per species always affected training success, the greater the imbalance the greater the effect. However, this could be largely compensated for by adjusting the networks using a posteriori probabilities, estimated as network output values.     

Learning the local landscape of protein structures with convolutional neural networks

One fundamental problem of protein biochemistry is to predict protein structure from amino acid sequence. The inverse problem, predicting either entire sequences or individual mutations that are consistent with a given protein structure, has received much less attention even though it has important applications in both protein engineering and evolutionary biology. Here, we ask whether 3D convolutional neural networks (3D CNNs) can learn the local fitness landscape of protein structure to reliably predict either the wild-type amino acid or the consensus in a multiple sequence alignment from the local structural context surrounding site of interest. We find that the network can predict wild type with good accuracy, and that network confidence is a reliable measure of whether a given prediction is likely going to be correct or not. Predictions of consensus are less accurate and are primarily driven by whether or not the consensus matches the wild type. Our work suggests that high-confidence mis-predictions of the wild type may identify sites that are primed for mutation and likely targets for protein engineering.

     

A new neural network architecture with associative memory, pruning and order-sensitive learning

A new paradigm of neural network architecture is proposed that works as associative memory along with capabilities of pruning and order-sensitive learning. The network has a composite structure wherein each node of the network is a Hopfield network by itself. The Hopfield network employs an order-sensitive learning technique and converges to user-specified stable states without having any spurious states. This is based on geometrical structure of the network and of the energy function. The network is so designed that it allows pruning in binary order as it progressively carries out associative memory retrieval. The capacity of the network is 2n, where n is the number of basic nodes in the network. The capabilities of the network are demonstrated by experimenting on three different application areas, namely a Library Database, a Protein Structure Database and Natural Language Understanding.