James P. Coughlin

A neural network for target classification using passive sonar

This concerns the design, tratilng, test and evaluation of a feed-forward neural network for classifying acoustic signals emitted by ships in transit by an omnidirectional hydrophore. Relatively noisy surface ships, moving rapidly at medium to long range, emit signals which superficially resemble those of quieter submarines, moving more slowly and closer to the listening device. The neural network approach is motivated by an obvious analogy to the sonar classifier of German and Sejnowski, who trained a neural network to classify active sonar returns from two undersea objects. The present problem can be solved by a similar network architecture, the outputs indicating which target type (if any) is present. The inputs represent the evolution of spectral densities for each of a number of time lags. Yet the number of target types and encounter geometries is far greater than could possibly be covered in any representative way by a training set comprised of real world data, Thus the task is to connect the network to a high fidelity, model-based digital simulator and to show that, by training on the output of the simulator, the neural network can learn to pass realistic tests. Permission to copy without fee all or part of this material is gmrtted provided that the copies are not made or distributed for direct commercial advantage, the ACM copyright notice and the title of the pubtieation and its date appear, and notice is given that copying is by permission of the Association for Computing Mach~:ry. TOCOPY otherwise,or to republish requires a fee andlor spectilc perrms sion. This describes a neural network design-and-testing exercise based on a simplistic model that captures a few of the salient features of the problem.

Time series prediction with linear and nonlinear adaptive networks

Backpropagation networks with a single hidden layer were trained to perform one-step prediction on a variety of scalar time series. The performance of such nets typically equals or exceeds that of the linear adaptive predictor of the same order. Comparisons of the linear and nonlinear predictors were made with periodic, chaotic, and random time series, including broadband ocean acoustic ambient noise.<<ETX>>

Optimization in cascaded Boltzmann machines with a temperature gradient: an alternative to simulated annealing

Boltzmann machines can be series-coupled with one-way retinotopic connections to produce good (suboptimal) solutions to classic optimization problems. Each isothermal Bokzmann machine in the series has a lower temperature than the one which drives it through excitatory spanning links. This scheme is an alternative to the usual simulated annealing approach in which a single network is cooled over the course of time. The simplest case features two identical subnets, one at a positive temperature and the other at zero temperature. Its performance in solving n-to-n assignment probems is compared to that obtained by simulated annealing with a geometric cooling schedule.

Stochastic models of information obsolescence

The stochastic modelling of information obsolescence is reviewed from a systematic and historical perspective. Burrells model of the use history of the individual data item is tested statistically against the citation histories of some physics articles. Alternative modelling approaches and their implications for the design of full-text electronic databases are considered.

Measures of serial data compressibility by neural network predictors

A time series or univariate random process is compressible if it is predictable. Experiments with a variety of processes readily show that adaptive neural networks are at least as effective as their linear counterparts in one-step-ahead prediction. The relationship between the predictive accuracy attained by the network, in the long run, and the closeness with which it can fit (and overfit) small segments of the same series in the course of many passes through the same data is examined. The findings suggest that the predictability of a process can be estimated by measuring the ease with which its increments can be overfitted.<<ETX>>

Calculating the information content of a temporally modulated spike train

An expression for the information content of a temporally modulated spike train is derived. With reference to published studies of the temporal encoding of visual stimuli by individual neurons, the expression suggests an information capacity of more than 5 bit/s. This result is insensitive to overspecification of the number of stimuli that the neuron is presumed able to resolve

Nash equilibria in multineuron games

An isomorphism is constructed between a neural network and a simple game. The equilibrium results of Nash can then be applied to the neural network. In particular, if the connecting matrix of the network is symmetric, the equilibrium can be found as the solution to an optimization problem.

An Alternative Derivation of the Hazard Rate

This paper suggests a simpler way of deriving the hazard rate equation from a differential equation of the Bernoulli type rather than through a probability argument. Such a derivation is often easier for engineering students to follow.

Comments on the parameter identificato in of a population model of china

Abstract The parametric model used by Zhu and Wan for expressing survival rate as a function of age can be replaced by a simpler model without discernable loss of accuracy. The parameters of this simpler model are subject to easy interpretation.