Optimization in an Error Backpropagation Neural Network Environment with a Performance Test on a Spectral Pattern Classification Problem

This paper attempts to develop a mathematically rigid framework for minimizing the cross-entropy function in an error backpropagating framework. In doing so, we derive the backpropagation formulae for evaluating the partial derivatives in a computationally efficient way. Various techniques of optimizing the multiple-class cross-entropy error function to train single hidden layer neural network classifiers with softmax output transfer functions are investigated on a real-world multispectral pixel-by-pixel classification problem that is of fundamental importance in remote sensing. These techniques include epoch-based and batch versions of backpropagation of gradient descent, PR-conjugate gradient, and BPGS quasi-Newton errors. The method of choice depends upon the nature of the learning task and whether one wants to optimize learning for speed or classification performance. It was found that, comparatively considered, gradient descent error backpropagation provided the best and most stable out-of-sample performance results across batch and epoch-based modes of operation. If the goal is to maximize learning speed and a sacrifice in classification accuracy is acceptable, then PR-conjugate gradient error backpropagation tends to be superior. If the training set is very large, stochastic epoch-based versions of local optimizers should be chosen utilizing a larger rather than a smaller epoch size to avoid unacceptable instabilities in the classification results.     

The Return of Zipf: Towards a Further Understanding of the Rank-Size Distribution

We offer a general-equilibrium economic approach to Zipf's Law or, more generally, the rank-size distribution—the striking empirical regularity concerning the size distribution of cities. We provide some further understanding of Zipf's Law by incorporating negative feedbacks (congestion) in a popular model of economic geography and international trade. This model allows the powers of agglomeration and spreading to be in long-run equilibrium, which enhances our understanding of the existence of a rank-size distribution of cities.     

Gendering resistance: British colonial narratives of wartime New Zealand

In 1850s New Zealand, the British Crown forced the purchase of indigenous Maori land at Taranaki, an event that hastened the brutal land wars of the 1860s. A group of British church and government officials and their wives voiced their opposition to this land confiscation, participating in what became known as the «pamphlet war» over Taranaki. The authors compare the anti-colonial writings of a number of prominent colonial men and women, drawing out the significance of both ideological and embodied gendered differences and the discursive representations of place produced out of them. In addition to differences and similarities in genre, content, and audience, the authors highlight the significance—and spatiality—of anti-colonial Maori testimony that appeared in one of the pivotal texts.     

A Supermarket of anthropology: AAA 2002

Conference reviewed:
Mette Louise Berg, A Supermarket of Anthropology AAA Meeting, New Orleans, 20–24 November 2002     

Estimating the Marginal Willingness to Pay for Commuting

With informational frictions on the labor market, hedonic wage regressions provide biased estimates of the willingness to pay for job attributes. We show that a recent theoretical result, which states that the variation in job durations provides a basis for obtaining good estimates, can be generalized to a wide class of search models. We apply this result by estimating the marginal willingness of employed workers to pay for commuting, using Dutch longitudinal data. The average willingness to pay for one hour commuting is estimated to equal almost half of the hourly wage rate.     

Optimization in an Error Backpropagation Neural Network Environment with a Performance Test on a Spectral Pattern Classification Problem

This paper attempts to develop a mathematically rigid framework for minimizing the cross-entropy function in an error backpropagating framework. In doing so, we derive the backpropagation formulae for evaluating the partial derivatives in a computationally efficient way. Various techniques of optimizing the multiple-class cross-entropy error function to train single hidden layer neural network classifiers with softmax output transfer functions are investigated on a real-world multispectral pixel-by-pixel classification problem that is of fundamental importance in remote sensing. These techniques include epoch-based and batch versions of backpropagation of gradient descent, PR-conjugate gradient, and BFGS quasi-Newton errors. The method of choice depends upon the nature of the learning task and whether one wants to optimize learning for speed or classification performance. It was found that, comparatively considered, gradient descent error backpropagation provided the best and most stable out-of-sample performance results across batch and epoch-based modes of operation. If the goal is to maximize learning speed and a sacrifice in classification accuracy is acceptable, then PR-conjugate gradient error backpropagation tends to be superior. If the training set is very large, stochastic epoch-based versions of local optimizers should be chosen utilizing a larger rather than a smaller epoch size to avoid unacceptable instabilities in the classification results.     

Infrared radiofluorescence (IR-RF) dating of middle pleistocene fluvial archives of the Heidelberg Basin (Southwest Germany)

The infrared radiofluorescence (IR-RF) dating technique was applied to eight fluvial samples that were collected from two sediment cores at the Heidelberg Basin located near Viernheim and Ludwigshafen in southwest Germany. Based on the IR-RF derived ages of the samples it was possible to establish a chronological framework for the Mid-Pleistocene fluvial deposits of the Heidelberg Basin. The results allow us to distinguish between four main periods of aggradation. The lowermost sample taken from 100 m core depth lead to an IR-RF age of 643 ± 28 ka pointing to a Cromerian period of aggradation (OIS 17–16). For the Elsterian it is now possible to distinguish between two aggradation periods, one occurring during the Lower Elsterian period (OIS 15) and a second during the Upper Elsterian period (OIS 12–11). For the so called Upper interlayer (or “Oberer Zwischenhorizont” — a layer of organic-rich and finer-grained deposits), the IR-RF results point to a deposition age of around 300 ka, with samples taken directly on top and out of this layer yielding IR-RF ages of 288 ± 19 ka and 302 ± 19 ka, respectively. Hence, the measured IR-RF ages clearly point to a deposition during the Lower Saalian period (OIS 9–8) whereas earlier studies assumed a Cromerian age for the sediments of the Upper Interlayer based on pollen records and also mollusc fauna. The new IR-RF dataset indicates that significant hiatuses are present within the fluvial sediment successions. In particular the Eemian and Upper Saalian deposits are missing in this part of the northern Upper Rhine Graben, as the 300 ka deposits are directly overlain by Weichselian fluvial sediments. It is obvious that time periods of increased fluvial aggradation were interrupted by time periods of almost no aggradation or erosion which should have been mainly triggered by phases of increased and decreased subsidence of the Heidelberg Basin.