A survey of partially connected neural networks

Almost all artificial neural networks are by default fully connected, which often implies a high redundancy and complexity. Little research has been devoted to the study of partially connected neural networks, despite its potential advantages like reduced training and recall time, improved generalization capabilities, reduced hardware requirements, as well as being a step closer to biological reality. This publication presents an extensive survey of the various kinds of partially connected neural networks, clustered into a clear framework, followed by a detailed comparative discussion.     

Everything is connected: Graph neural networks

In many ways, graphs are the main modality of data we receive from nature. This is due to the fact that most of the patterns we see, both in natural and artificial systems, are elegantly representable using the language of graph structures. Prominent examples include molecules (represented as graphs of atoms and bonds), social networks and transportation networks. This potential has already been seen by key scientific and industrial groups, with already-impacted application areas including traffic forecasting, drug discovery, social network analysis and recommender systems. Further, some of the most successful domains of application for machine learning in previous years—images, text and speech processing—can be seen as special cases of graph representation learning, and consequently there has been significant exchange of information between these areas. The main aim of this short survey is to enable the reader to assimilate the key concepts in the area, and position graph representation learning in a proper context with related fields.     

Pruning artificial neural networks using neural complexity measures

This paper describes a new method for pruning artificial neural networks, using a measure of the neural complexity of the neural network. This measure is used to determine the connections that should be pruned. The measure computes the information-theoretic complexity of a neural network, which is similar to, yet different from previous research on pruning. The method proposed here shows how overly large and complex networks can be reduced in size, whilst retaining learnt behaviour and fitness. The technique proposed here helps to discover a network topology that matches the complexity of the problem it is meant to solve. This novel pruning technique is tested in a robot control domain, simulating a racecar. It is shown, that the proposed pruning method is a significant improvement over the most commonly used pruning method Magnitude Based Pruning. Furthermore, some of the pruned networks prove to be faster learners than the benchmark network that they originate from. This means that this pruning method can also help to unleash hidden potential in a network, because the learning time decreases substantially for a pruned a network, due to the reduction of dimensionality of the network.     

String matching artificial neural networks.

Three artificial neural networks (ANNs) are proposed for solving a variety of on- and off-line string matching problems. The ANN structure employed as the building block of these ANNs is derived from the harmony theory (HT) ANN, whereby the resulting string matching ANNs are characterized by fast match-mismatch decisions, low computational complexity, and activation values of the ANN output nodes that can be used as indicators of substitution, insertion (addition) and deletion spelling errors.     

Cardiovascular control with artificial neural networks

This paper describes the use of artificial neural networks to model cardiovascular autonomic control in a study of the hemodynamic changes associated with space flight. Cardiovascular system models were created including four parameters: heart rate, contractility, peripheral resistance, and venous tone. Artificial neural networks were then designed and trained. A technique known as backpropagation networking was used and the results of the application of this technique to heart rate control are presented and discussed.     

Bioprocess model building using artificial neural networks

Artificial neural networks are made upon of highly interconnected layers of simple neuron-like nodes. The neurons act as non-linear processing elements within the network. An attractive property of artificial neural networks is that given the appropriate network topology, they are capable of learning and characterising non-linear functional relationships. Furthermore, the structure of the resulting neural network based process model may be considered generic, in the sense that little prior process knowledge is required in its determination. The methodology therefore provides a cost efficient and reliable process modelling technique. One area where such a technique could be useful is biotechnological systems. Here, for example, the use of a process model within an estimation scheme has long been considered an effective means of overcoming inherent on-line measurement problems. However, the development of an accurate process model is extremely time consuming and often results in a model of limited applicability. Artificial neural networks could therefore prove to be a useful model building tool when striving to improve bioprocess operability. Two large scale industrial fermentation systems have been considered as test cases; a fed-batch penicillin fermentation and a continuous mycelial fermentation. Both systems serve to demonstrate the utility, flexibility and potential of the artificial neural network approach to process modelling.     

Modeling of pain using artificial neural networks

In dealing with human nervous system, the sensation of pain is as sophisticated as other physiological phenomena. To obtain an acceptable model of the pain, physiology of the pain has been analysed in the present paper. Pain mechanisms are explained in block diagram representation form. Because of the nonlinear interactions existing among different sections in the diagram, artificial neural networks (ANNs) have been exploited. The basic patterns associated with chronic and acute pain have been collected and then used to obtain proper features for training the neural networks. Both static and dynamic representations of the ANNs were used in this regard. The trained networks then were employed to predict response of the body when it is exposed to special excitations. These excitations have not been used in the training phase and their behavior is interesting from the physiological view. Some of these predictions can be inferred from clinical experimentations. However, more clinical tests have to be accomplished for some of the predictions.     

Microarray-based cancer diagnosis with artificial neural networks

In recent years, the advent of experimental methods to probe gene expression profiles of cancer on a genome-wide scale has led to widespread use of supervised machine learning algorithms to characterize these profiles. The main applications of these analysis methods range from assigning functional classes of previously uncharacterized genes to classification and prediction of different cancer tissues. This article surveys the application of machine learning algorithms to classification and diagnosis of cancer based on expression profiles. To exemplify the important issues of the classification procedure, the emphasis of this article is on one such method, namely artificial neural networks. In addition, methods to extract genes that are important for the performance of a classifier, as well as the influence of sample selection on prediction results are discussed.     

A dynamical behaviour of active regions in randomly connected neural networks

The field of the randomly connected neural network is approximately formulated by Griffith's equation, regarding the network as being continuous. An integral representation of Griffith's equation is derived. If a relative refractory period can be ignored, it is $\text{X(x,t)}\text{=1}\text{∫}\text{o}\text{∞}\text{ds}\text{∫}\text{?vs}\text{vs}\text{dn}\text{kv}\text{2t}\text{e}{}^{\mathrm{?avs}}\text{X(x?n, t?s) ? θ}$ where X(x, t) corresponds to the firing rate and θ means the threshold of the neural firing, τ the absolute refractory period and v the velocity for the spike potential travelling down the axon. The above equation is formally analogous to Caianiello's equation, but the former describes the more macroscopic behaviour of the neural network than the latter. With the aid of computer simulation, appropriate solutions are successfully obtained.In regions where X = 1, neurones are firing at a high constant rate of $\text{1}\text{τ}$ (active regions). In regions where X = 0, there is no firing of neurones (resting regions). In the neural net for which $\text{0 <}\text{a}{}^2\text{τθ}\text{k}\text{< 1}$, the net is generally a mixture of the active regions and of the resting regions. In the case that a large active region is in contact with a large resting region, the propagation velocity of boundary between the two regions tends to the velocity u given by $\text{u = (1 ?}\text{2a}{}^2\text{τθ}\text{k}\text{)v}$. This expression of velocity u was deduced from the fact that there exists a solution of the type X(x, t) = 1 (ut ? x) for equation (A). In the case of $\text{0 <}\text{a}{}^2\text{τθ}\text{k}\text{< 0 · 5}$, the active region grows and in the case of $\text{0 · 5 <}\text{a}{}^2\text{τθ}\text{k}\text{< 1}$, the resting region grows. A fatigue effect is introduced, for which it is hard for neurones to maintain firing states. In this case an active region of definite width L propagates with constant velocity u′. The dependence of L and u′ on characteristics of neural network and on the fatigue effect is investigated.     

Applications of artificial neural networks predicting macroinvertebrates in freshwaters

To facilitate decision support in freshwater ecosystem protection and restoration management, habitat suitability models can be very valuable. Data driven methods such as artificial neural networks (ANNs) are particularly useful in this context, seen their time-efficient development and relatively high reliability. However, specialized and technical literature on neural network modelling offers a variety of model development criteria to select model architecture, training procedure, etc. This may lead to confusion among ecosystem modellers and managers regarding the optimal training and validation methodology. This paper focuses on the analysis of ANN development and application for predicting macroinvertebrate communities, a species group commonly used in freshwater assessment worldwide. This review reflects on the different aspects regarding model development and application based on a selection of 26 papers reporting the use of ANN models for the prediction of macroinvertebrates. This analysis revealed that the applied model training and validation methodologies can often be improved and moreover crucial steps in the modelling process are often poorly documented. Therefore, suggestions to improve model development, assessment and application in ecological river management are presented. In particular, data pre-processing determines to a high extent the reliability of the induced models and their predictive relevance. This also counts for the validation criteria, that need to be better tuned to the practical simulation requirements. Moreover, the use of sensitivity methods can help to extract knowledge on the habitat preference of species and allow peer-review by ecological experts. The selection of relevant input variables remains a critical challenge as well. Model coupling is a missing crucial step to link human activities, hydrology, physical habitat conditions, water quality and ecosystem status. This last aspect is probably the most valuable aspect to enable decision support in water management based on ANN models.     

Measuring context dependency in birdsong using artificial neural networks

Context dependency is a key feature in sequential structures of human language, which requires reference between words far apart in the produced sequence. Assessing how long the past context has an effect on the current status provides crucial information to understand the mechanism for complex sequential behaviors. Birdsongs serve as a representative model for studying the context dependency in sequential signals produced by non-human animals, while previous reports were upper-bounded by methodological limitations. Here, we newly estimated the context dependency in birdsongs in a more scalable way using a modern neural-network-based language model whose accessible context length is sufficiently long. The detected context dependency was beyond the order of traditional Markovian models of birdsong, but was consistent with previous experimental investigations. We also studied the relation between the assumed/auto-detected vocabulary size of birdsong (i.e., fine- vs. coarse-grained syllable classifications) and the context dependency. It turned out that the larger vocabulary (or the more fine-grained classification) is assumed, the shorter context dependency is detected.     

Single cortical neurons as deep artificial neural networks

1. Download : Download high-res image (228KB)
2. Download : Download full-size image

     

Homeostatic plasticity improves signal propagation in continuous-time recurrent neural networks

Continuous-time recurrent neural networks (CTRNNs) are potentially an excellent substrate for the generation of adaptive behaviour in artificial autonomous agents. However, node saturation effects in these networks can leave them insensitive to input and stop signals from propagating. Node saturation is related to the problems of hyper-excitation and quiescence in biological nervous systems, which are thought to be avoided through the existence of homeostatic plastic mechanisms. Analogous mechanisms are here implemented in a variety of CTRNN architectures and are shown to increase node sensitivity and improve signal propagation, with implications for robotics. These results lend support to the view that homeostatic plasticity may prevent quiescence and hyper-excitation in biological nervous systems.     

Prediction of fingers posture using artificial neural networks

Predicting the hand and fingers posture during grasping tasks is an important issue in the frame of biomechanics. In this paper, a technique based on neural networks is proposed to learn the inverse kinematics mapping between the fingertip 3D position and the corresponding joint angles. Finger movements are obtained by an instrumented glove and are mapped to a multichain model of the hand. From the fingertip desired position, the neural networks allow predicting the corresponding finger joint angles keeping the specific subject coordination patterns. Two sets of movements are considered in this study. The first one, the training set, consisting of free fingers movements is used to construct the mapping between fingertip position and joint angles. The second one, constructed for testing purposes, is composed of a sequence of grasping tasks of everyday-life objects. The maximal mean error between fingertip measured position and fingertip position obtained from simulated joint angles and forward kinematics is 0.99+/-0.76mm for the training set and 1.49+/-1.62mm for the test set. Also, the maximal RMS error of joint angles prediction is 2.85 degrees and 5.10 degrees for the training and test sets respectively, while the maximal mean joint angles prediction error is -0.11+/-4.34 degrees and -2.52+/-6.71 degrees for the training and test sets, respectively. Results relative to the learning and generalization capabilities of this architecture are also presented and discussed.     

Clustering proteins into families using artificial neural networks

An incorrect version of Figure 3 was published in the abovearticle, the corrected version is reproduced below.     

Creating artificial human genomes using generative neural networks

Generative models have shown breakthroughs in a wide spectrum of domains due to recent advancements in machine learning algorithms and increased computational power. Despite these impressive achievements, the ability of generative models to create realistic synthetic data is still under-exploited in genetics and absent from population genetics. Yet a known limitation in the field is the reduced access to many genetic databases due to concerns about violations of individual privacy, although they would provide a rich resource for data mining and integration towards advancing genetic studies. In this study, we demonstrated that deep generative adversarial networks (GANs) and restricted Boltzmann machines (RBMs) can be trained to learn the complex distributions of real genomic datasets and generate novel high-quality artificial genomes (AGs) with none to little privacy loss. We show that our generated AGs replicate characteristics of the source dataset such as allele frequencies, linkage disequilibrium, pairwise haplotype distances and population structure. Moreover, they can also inherit complex features such as signals of selection. To illustrate the promising outcomes of our method, we showed that imputation quality for low frequency alleles can be improved by data augmentation to reference panels with AGs and that the RBM latent space provides a relevant encoding of the data, hence allowing further exploration of the reference dataset and features for solving supervised tasks. Generative models and AGs have the potential to become valuable assets in genetic studies by providing a rich yet compact representation of existing genomes and high-quality, easy-access and anonymous alternatives for private databases.     

Sleep classification in infants based on artificial neural networks.

The study reports on the possibility of classifying sleep stages in infants using an artificial neural network. The polygraphic data from 4 babies aged 6 weeks, 6 months and 1 year recorded over 8 hours were available for classification. From each baby 22 signals were recorded, digitized and stored on an optical disc. Subsets of these signals and additional calculated parameters were used to obtain data vectors, each of which represents an interval of 30 sec. For classification, two types of neural networks were used, a Multilayer Perceptron and a Learning Vector Quantizer. The teaching input for both networks was provided by a human expert. For the 6 sleep classes in babies aged 6 months, a 65% to 80% rate of correct classification (4 babies) was obtained for the testing data not previously seen.