Number Of Output Units Research Articles

Stabilizing patterns in time: Neural network approach.

Recurrent and feedback networks are capable of holding dynamic memories. Nonetheless, training a network for that task is challenging. In order to do so, one should face non-linear propagation of errors in the system. Small deviations from the desired dynamics due to error or inherent noise might have a dramatic effect in the future. A method to cope with these difficulties is thus needed. In this work we focus on recurrent networks with linear activation functions and binary output unit. We characterize its ability to reproduce a temporal sequence of actions over its output unit. We suggest casting the temporal learning problem to a perceptron problem. In the discrete case a finite margin appears, providing the network, to some extent, robustness to noise, for which it performs perfectly (i.e. producing a desired sequence for an arbitrary number of cycles flawlessly). In the continuous case the margin approaches zero when the output unit changes its state, hence the network is only able to reproduce the sequence with slight jitters. Numerical simulation suggest that in the discrete time case, the longest sequence that can be learned scales, at best, as square root of the network size. A dramatic effect occurs when learning several short sequences in parallel, that is, their total length substantially exceeds the length of the longest single sequence the network can learn. This model easily generalizes to an arbitrary number of output units, which boost its performance. This effect is demonstrated by considering two practical examples for sequence learning. This work suggests a way to overcome stability problems for training recurrent networks and further quantifies the performance of a network under the specific learning scheme.

Measuring the Efficiency of Chilean Primary Healthcare Centres

This paper examines whether different approaches for estimating the technical efficiency of a primary healthcare centre —PHC— can have significantly different results. Although effective regulation demands the development of measures for establishing good operation levels, few studies have compared results from a data envelopment analysis —DEA— and stochastic frontier analysis —SFA—in a PHC. Furthermore, to the best of our knowledge, no studies have used Chilean nationwide data. Efficiency refers to the relation between the number of output units per input unit. Effectiveness evaluates the outcome of medical care and can be influenced by efficiency.The data that were used in this study came from 259 Chilean municipalities. We included two outputs – medical and check-up visits – and three inputs – staff, general service and drugs expenses. For the DEA, we used a variable return to scale output-oriented model. Before applying the SFA, a principal component analysis —PCA— combined the two outputs. The results were similar for both methods: the SFA efficiency averages were 70.89% and 65.83% and the DEA averages were 68.37% and 54.46% for the urban and rural municipalities, respectively. In addition to defining a frontier of “best practice”, this study merged the PCA with the SFA to form an innovative approach to combining outputs.

On the use of self-organizing maps for clustering and visualization

We show that the number of output units used in a self-organizing map (SOM) influences its applicability for either clustering or visualization. By reviewing the appropriate literature and theory and own empirical results, we demonstrate that SOMs can be used for clustering or visualization separately, for simultaneous clustering and visualization, and even for clustering via visualization. For all these different kinds of application, SOM is compared to other statistical approaches. This will show SOM to be a flexible tool which can be used for various forms of explorative data analysis but it will also be made obvious that this flexibility comes with a price in terms of impaired performance.

Analyzing a Combination of Factors for Thinning Trees with a Neural Network

A neural network, which is a computer model applied to an artificially simulated process of neurons, has various advantages. For example, it allows treating both quantitative and qualitative data at the same time. In a previous paper, the author suggested that a neural network was an effective method to analyze subjective forest information such as the selection of trees for thinning. The purpose of this paper is to analyze the relationship between a combination of factors and degrees of contribution to a neural network structure. The AIC (Akaike's information criterion) was used as an information criterion for model selection and a three-layer backpropagation model was used as the learning algorithm. It was found that: (1) qualitative factors (i.e., crown volume, stem quality and “defects”) conferred more accurate estimation results than quantitative factors (DBH and height), and (2) “defects” was the most important factor, especially if the number of output units was two, because it was possible to discriminate sufficiently even in the absence of other factors. It is important to consider a combination of factors in the model when applying a neural network; Additionally, few factors usually result in more useful and meaningful models.

Diagrammatic Approach to Capacity-Constrained Price Discrimination

This paper presents a diagrammatic solution to the firm’s profit-maximizing price discrimination problem in the face of capacity constraints. Airlines, hotels, and other firms practice yield management, allocating fixed capacity to customer groups paying different prices. In these cases, the firm’s short-run problem is not a decision about production levels, but it is one of allocating a fixed number of output units among customers. Our diagram shows that the conditions for profit-maximizing price discrimination are very different under these circumstances than in the standard model in which the firm is not constrained by capacity.

A neural network approach to pulse radar detection

A multilayer feedforward neural network is applied to pulse compression. The 13-element Barker code and the maximum-length sequences (m-sequences) with lengths 15, 31, and 63 b were used as the signal codes, and four networks were implemented, respectively. In each of these networks, the number of input units was the same as the signal length while the number of hidden units was three and the number of output units was one. In training each of these networks, backpropagation learning was used and the number of training epochs was 500. Using this approach, a more than 40 dB output peak signal-to-sidelobe ratio can be achieved. These fault-tolerant neural networks can provide a robust means for pulse radar detection. >

Mean Field Theory Neural Networks for Feature Recognition, Content Addressable Memory and Optimization

Abstract Various applications of the mean field theory (MFT) technique for obtaining solutions close to optimal minima in feedback networks are reviewed. Using this method in the context of the Boltzmann machine gives rise to a fast deterministic learning algorithm with a performance comparable with that of the backpropagation algorithm (BP) in feature recognition applications. Since MFT learning is bidirectional its use can be extended from purely functional mappings to a content addressable memory. The storage capacity of such a network grows like O (10–20)nH with the number of hidden units. The MFT learning algorithm is local and thus it has an advantage over BP with respect to VLSI implementations. It is also demonstrated how MFT and BP are related in situations where the number of input units is much larger than the number of output units. In the context of-finding good solutions to difficult optimization problems the MFT technique again turns out to be extremely powerful. The quality of the solution...

Pattern-recognition by an artificial network derived from biologic neuronal systems

A novel artificial neural network, derived from neurobiological observations, is described and examples of its performance are presented. This DYnamically STable Associative Learning (DYSTAL) network associatively learns both correlations and anticorrelations, and can be configured to classify or restore patterns with only a change in the number of output units. DYSTAL exhibits some particularly desirable properties: computational effort scales linearly with the number of connections, i.e., it is O(N) in complexity; performance of the network is stable with respect to network parameters over wide ranges of their values and over the size of the input field; storage of a very large number of patterns is possible; patterns need not be orthogonal; network connections are not restricted to multi-layer feed-forward or any other specific structure; and, for a known set of deterministic input patterns, the network weights can be computed, a priori, in closed form. The network has been associatively trained to perform the XOR function as well as other classification tasks. The network has also been trained to restore patterns obscured by binary or analog noise. Neither global nor local feedback connections are required during learning; hence the network is particularly suitable for hardware (VLSI) implementation.