No abstract

Concerns the problem of finding weights for feed-forward networks in which threshold functions replace the more common logistic node output function. The advantage of such weights is that the complexity of the hardware implementation of such networks is greatly reduced. If the task to be learned does not change over time, it may be sufficient to find the correct weights for a threshold function network off-line and to transfer these weights to the hardware implementation. This paper provides a mathematical foundation for training a network with standard logistic function nodes and gradually altering the function to allow a mapping to a threshold unit network. The procedure is analogous to taking the limit of the logistic function as the gain parameter goes to infinity. It is demonstrated that, if the error in a trained network is small, a small change in the gain parameter will cause a small change in the network error. The result is that a network that must be implemented with threshold functions can first be trained using a traditional back propagation network using gradient descent, and further trained with progressively steeper logistic functions. In theory, this process could require many repetitions. In simulations, however, the weights have be successfully mapped to a true threshold network after a modest number of slope changes. It is important to emphasize that this method is only applicable to situations for which off-line learning is appropriate.

Abstract. The accuracy and efficiency of four approaches to identifying clouds and aerosols in remote sensing imagery are compared. These approaches are as follows: a maximum likelihood classifier, a paired histogram technique, a hybrid class elimination approach, and a backpropagation neural network. Regional comparisons were conducted on advanced very high resolution radiometer (AVHRR) local area coverage (LAC) scenes from the polar regions, desert areas, and regions of biomass-burning, areas which are known to be particularly difficult. For the polar, desert, and biomass burning regions, the maximum likelihood classifier achieved 94-97% accuracy, the neural network achieved 95-96% accuracy, and the paired histogram approach achieved 93-94% accuracy. The primary advantage to the class elimination scheme lies in its speed; its accuracy of 94-96% is comparable to that of the maximum likelihood classifier. Experiments also clearly demonstrate the effectiveness of decomposing a single global classifier into separate regional classifiers, since the regional classifiers can be more finely tuned to recognize local conditions. In addition, the effectiveness of using composite features is compared to the simpler approach of using the five AVHRR channels and the reflectance of channel 3 treated as a sixth channel as the elements of the feature vector. The results varied, demonstrating that the features cannot be chosen independently of the classifier to be used. It is also shown that superior results can obtained by training the classifiers using subclass information and collapsing the subclasses after classification. Finally, ancillary data were incorporated into the classifiers, consisting of a land/water mask, a terrain map, and a computed sunglint probability. While the neural network did not benefit from this information, the accuracy of the maximum likelihood classifier improved by 1%, and the accuracy of the paired histogram method increased by up to 4%.

No abstract

We describe a massively parallel implementation of genetic algorithms using a MasPar MP-1 data-parallel computer. Modification of the sequential genetic algorithm required that several important issues be addressed, in particular how to implement the selection operator with a minimum of inter-processor communication. A speed up of at least 145 times relative to a comparable sequential algorithm running on a Sun SparcStation was achieved. The performance of the MasPar is even more impressive when populations are too large to tit into the physical memory of the Spare.