Traditional neural network

Catasus et al. [67] studied two types of neural networks traditional multilayer perceptron neural networks and generalised regression neural networks (GRNNs) to correct for nonlinear matrix effects and long-term signal drift in ICP-AES. 

Predictive methods aimed at extracting secondary structural information from the linear primary sequence make extensive use of neural networks, traditionally used for analysis of patterns and trends. Basically, a neural network provides computational processes the ability to learn in an attempt to approximate human learning versus following instructions blindly in a sequential marmer. Every nemal network has an input layer and an output layer. In the case of secondary structure prediction, the input layer would be information from the sequence itself, and the output layer would be the probabilities of whether a particular residue could form a particular structure. Between the input and output layers would be one or more hidden layers where the actual learning would take place. This is accomplished by providing a training data set for the network. Here, an appropriate training set would be all sequences for which three-dimensional structures have been deduced. The network can process this information to look for what are possibly weak relationships between an amino acid sequence and the structures they can form in a particular context. A more complete discussion of neural networks as applied to secondary structure prediction can be found in Kneller et al. (1990). 

Numeric-to-numeric transformations are used as empirical mathematical models where the adaptive characteristics of neural networks learn to map between numeric sets of input-output data. In these modehng apphcations, neural networks are used as an alternative to traditional data regression schemes based on regression of plant data. Backpropagation networks have been widely used for this purpose. 

Especially the last few years, the number of applications of neural networks has grown exponentially. One reason for this is undoubtedly the fact that neural networks outperform in many applications the traditional (linear) techniques. The large number of samples that are needed to train neural networks remains certainly a serious bottleneck. The validation of the results is a further issue for concern. 

With increasing toxicity data of various kinds, more rehable predictions based on structure-toxicity relationships of toxic endpoints can be attempted [31-36]. Even the Internet can be used as a source for toxicity data, albeit with caution [37]. A number of predictive methods have been compared from a regulatory perspective [35]. Often traditional QSAR approaches using multiple Hnear regression are used [38]. Newer approaches include the use of neural networks in structure-toxicity relationships... 

Key Words 2D-QSAR traditional QSAR 3D-QSAR nD-QSAR 4D-QSAR receptor-independent QSAR receptor-dependent QSAR high throughput screening alignment conformation chemometrics principal components analysis partial least squares artificial neural networks support vector machines Binary-QSAR selecting QSAR descriptors. 

Traditional regression-type models have been linear and quadratic regression models. Linear and quadratic regression models unfortunately impose further constraints upon the nature of the process nonlinearity as such, these models are limited in the range of their applicability. A relatively new nonlinear regression-type model—the Artificial Neural Network (ANN)—is not as limited, and is worthy of additional discussion. 

There have been many books and reviews written on the subject of NN and parallel computing. Only a token one is listed here, for those who need a traditional book reference (Haykin, 1999). It will probably be obsolete before this book is published. Otherwise, a wealth of up-to-date information is always available on the Internet where a neural networks entry produces an avalanche of information. Both lead articles cited for Chapter 10 (Hierlemann et al 1996) and (Jurs et al., 2000) discuss their applications in the context of chemical and biological sensing. 

While most combinatorial researches reported up to now involve the use of GA, using the traditional crossover and mutation operators (e.g. WGS 1), it has also been proposed to design new operators for each specific application, to improve search efficiency by means of knowledge extraction [32]. Hence, new methods that combine ES with a knowledge extraction engine have been reported recently within the field of heterogeneous catalysis, such as mining association rules [12, 18, 30, 33] and neural networks [19, 29, 34]. 

In recent years, neuron models have been successfully used they are more stable than traditional climate models. For instance, Pasini et al. (2006) considered an application of the neural network for climate modeling. The study was carried out into the temperature trend on regional and global scales for the last 140 years. It showed that the model based on the neural network reproduces with high accuracy the non-linear effects observed in temperature variations over the northern Atlantic. 

Not all computers carry out computations in a traditional way. Neural networks are another form of computer that receive input signals and produce output signals that characterize what was input. These computers can be taught to recognize complex patterns. For example, they can be shown a picture of a person and then can recognize another picture of the same person, even when viewed from a different perspective. They have also been taught to recognize connected speech. 

Artificial neural networks are versatile tools for a number of applications, including bioinformatics. However, they are not thinking machines nor are they black boxes to blindly feed data into with expectations of miraculous results. Neural networks are typically computer software implementations of algorithms, which fortunately may be represented by highly visual, often simple diagrams. Neural networks represent a powerful set of mathematical tools, usually highly nonlinear in nature, that can be used to perform a number of traditional statistical chores such as classification, pattern recognition and feature extraction. 

Bayesian neural networks (BNNs) are an alternative to the more traditional ANNs. The main advantage with BNNs is that they are less prone to overtraining compared to ANNs. BNNs are based on Bayesian probabilistics for the network training. Network weights are determined by Bayesian inference. BNNs have been successfully used together with automatic relevance determination (ARD) for the selection of relevant descriptors to model aqueous solubility [89]. For a good review on BNNs, see Ref. [90]. 

---