The convergence of backpropagation trained neural networks for various weight update frequencies

Fear sometimes returns after successful fear attenuation via in vivo exposure to fear signals. Post-treatment return of fear is of considerable interest both practically and theoretically, but factors associated with return of fear are poorly understood due to conflicting results from procedurally diverse experiments. This paper reports two very similar experiments in which fear of animal specimens was weakened then allowed to return so that factors associated with return of fear could be studied. In each experiment attentional focus versus distraction during exposure served as a between-subjects independent variable. In each case, attempts also were made to predict return of fear via several nonmanipulated variables: initial fear, initial avoidance during voluntary exposure, initial heart rate during voluntary exposure, and speed of fear reduction during repeated exposure trials. With the sample sizes used there was only suggestive evidence that return of fear was associated with distraction during exposure, and with relatively rapid fear decline during exposure. More importantly, the experiments are offered as standard, replicable models for research that will permit procedurally homogeneous investigations of variables with which return of fear is associated.     

Computer-assisted image classification: use of neural networks in anatomic pathology

Artificial neural networks (ANN) implemented on digital computers have received much attention for interpretation of images in pathology and cytology. Most such images are too complex for current ANN to interpret directly; instead, ANN usually classify the images according to numeric features extracted from them. In experiments on three distinct image classification problems, ANN classifiers performed as well or better than multivariate linear discriminant analysis (a traditional parametric statistical classifier). ANN empirically define non-linear multivariate decision boundaries, and can combine non-contiguous feature areas in mapping a classification. However, many training cases are required in order to map complex area boundaries precisely and with a low risk of 'overtraining.' Careful problem selection and attention to data dimensionality and format are important for efficient ANN use.     

Comparative evaluation of the use of artificial neural networks for modelling the epidemiology of schistosomiasis mansoni

There has been a marked increase in the application of approaches based on artificial intelligence (AI) in the field of computer science and medical diagnosis, but AI is still relatively unused in epidemiological settings. In this study we report results of the application of neural networks (NN; a special category of AI) to schistosomiasis. It was possible to design an NN structure which can process and fit epidemiological data collected from 251 schoolchildren in Egypt using the first year's data to predict second and third years' infection rates. Data collected over 3 years included age, gender, exposure to canal water and agricultural activities, medical history and examination, and stool and urine parasitology. Schistosoma mansoni infection rates were 50%, 78% and 66% at the baseline and the 2 follow-up periods, respectively. NN modelling was based on the standard back-propagation algorithm, in which we built a suitable configuration of the network, using the first year's data, that optimized performance. It was implemented on an IBM compatible computer using commercially available software. The performance of the NN model in the first year compared favourably with logistic regression (NN sensitivity = 83% (95% confidence interval [CI] 78-88%) and positive predictive value (PPV) = 63% (95% CI 57-69%); logistic regression sensitivity = 66% (95% CI 60%-72%) and PPV = 59% (95% CI 53%-65%). The NN model generalized the criteria for predicting infection over time better than logistic regression and showed more stability over time, as it retained its sensitivity and specificity and had better false positive and negative profiles than logistic regression. The potential of NN-based models to analyse and predict wide-scale control programme data, which are inevitably based on unstable egg excretion rates and insensitive laboratory techniques, is promising but still untapped.     

Synchrony in excitatory neural networks

Synchronization properties of fully connected networks of identical oscillatory neurons are studied, assuming purely excitatory interactions. We analyze their dependence on the time course of the synaptic interaction and on the response of the neurons to small depolarizations. Two types of responses are distinguished. In the first type, neurons always respond to small depolarization by advancing the next spike. In the second type, an excitatory postsynaptic potential (EPSP) received after the refractory period delays the firing of the next spike, while an EPSP received at a later time advances the firing. For these two types of responses we derive general conditions under which excitation destabilizes in-phase synchrony. We show that excitation is generally desynchronizing for neurons with a response of type I but can be synchronizing for responses of type II when the synaptic interactions are fast. These results are illustrated on three models of neurons: the Lapicque integrate-and-fire model, the model of Connor et al., and the Hodgkin-Huxley model. The latter exhibits a type II response, at variance with the first two models, that have type I responses. We then examine the consequences of these results for large networks, focusing on the states of partial coherence that emerge. Finally, we study the Lapicque model and the model of Connor et al. at large coupling and show that excitation can be desynchronizing even beyond the weak coupling regime.     

On the use of Bayesian methods for evaluating compartmental neural models

Computational modeling is being used increasingly in neuroscience. In deriving such models, inference issues such as model selection, model complexity, and model comparison must be addressed constantly. In this article we present briefly the Bayesian approach to inference. Under a simple set of commonsense axioms, there exists essentially a unique way of reasoning under uncertainty by assigning a degree of confidence to any hypothesis or model, given the available data and prior information. Such degrees of confidence must obey all the rules governing probabilities and can be updated accordingly as more data becomes available. While the Bayesian methodology can be applied to any type of model, as an example we outline its use for an important, and increasingly standard, class of models in computational neuroscience--compartmental models of single neurons. Inference issues are particularly relevant for these models: their parameter spaces are typically very large, neurophysiological and neuroanatomical data are still sparse, and probabilistic aspects are often ignored. As a tutorial, we demonstrate the Bayesian approach on a class of one-compartment models with varying numbers of conductances. We then apply Bayesian methods on a compartmental model of a real neuron to determine the optimal amount of noise to add to the model to give it a level of spike time variability comparable to that found in the real cell.     

A flexible algorithm for construction of 3-D vessel networks for use in thermal modeling

A new algorithm for the construction of artificial blood vessel networks is presented. The algorithm produces three-dimensional (3-D) geometrical representations of both arterial and venous networks. The key ingredient of the algorithm is a 3-D potential function defined in the tissue volume. This potential function controls the paths by which points are connected to existing vessels, thereby producing new vessel segments. The potential function has no physiological interpretation, but, by adjustment of parameters governing the potential, it is possible to produce networks that have physiologically meaningful geometrical properties. If desired, the veins can be generated counter current to the arteries. Furthermore, the potential function allows fashioning of the networks to the presence of bone or air cavities. The resulting networks can be used for thermal simulations of hyperthermia treatment.     

Artificial neural networks in cancer research

Current therapy does not cure the majority of patients with B cell non-Hodgkin's lymphoma (NHL) and further intensification does not benefit the patient. Therefore, new approaches are necessary. Immunotherapy has become again a major interest as a new treatment modality for B cell lymphoma since the discovery that the lymphoma specific Id can be presented to antigen-specific T cells. Vaccination of the tumour-bearing host is one of the major strategies to induce a T cell mediated anti-tumour immunity in vivo. For B cell lymphomas the lymphoma specific Id can be used as a tumour-specific antigen to stimulate T cells. Alternatively, the malignant B cells can be modified to become efficient antigen presenting cells (APCs) and present peptides from their own tumour-specific antigens to the autologous T cells. Currently explored and future vaccination strategies for B cell lymphoma will be discussed here.