Neural networks Kohonen

Figure 8-1J. Training ofa Kohonen neural network with a chirality code, The number of weights in a neuron is the same as the number of elements in the chirality code vector, When a chirality code is presented to the network, the neuron with the most similar weights to the chirality code is excited (this is the ivinning or central neuron) (see Section 9.5,3),...

X. H. Song and P.K. Hopke, Kohonen neural network as a pattern-recognition method, based on weight interpretation. Anal. Chim. Acta, 334 (1996) 57-66. 

Anzali, S., Bamickel, G., Krug, M., Sadowski, J., Wagener, M. and Gasteiger, J. (1996) Evaluation of molecular surface properties using a Kohonen neural network. In Neural Networks in QSAR and Design, Devillers, J. (Ed.), Academic Press, London. 

Some of the pioneering studies published by several reputed authors in the chemometrics field [55] employed Kohonen neural networks to diagnose calibration problems related to the use of AAS spectral lines. As they focused on classifying potential calibration lines, they used Kohonen neural networks to perform a sort of pattern recognition. Often Kohonen nets (which were outlined briefly in Section 5.4.1) are best suited to perform classification tasks, whereas error back-propagation feed-forwards (BPNs) are preferred for calibration purposes [56]. 

Balbinot et al. [36] classified Antarctic algae by applying Kohonen neural networks to a data set composed of 14 elements determined by ICP-OES. 

The utility of ANNs as a pattern recognition technique in the field of microbeam analysis was demonstrated by Ro and Linton [99]. Back-propagation neural networks were applied to laser microprobe mass spectra (LAMMS) to determine interparticle variations in molecular components. Selforganizing feature maps (Kohonen neural networks) were employed to extract information on molecular distributions within environmental microparticles imaged in cross-section using SIMS. 

L. Ralbinot, P. Smichowski, S. Farias, M. A. Z. Arruda, C. Vodopivez and R. J. Poppi, Classification of Antarctic algae by applying Kohonen neural network with 14 elements determined by inductively coupled plasma optical emission spectrometry, Spectrochim. Acta, Part B, 60(5), 2005, 725-730. 

Y. Vander Heyden, P. Vankeerberghen, M. Novic, J. Zupan and D. L. Massart, The application of Kohonen neural networks to diagnose calibration problems in atomic absorption spectrometry, Talanta, 51(3), 2000, 455-466. 

Another approach for solving the problem of representing data points in an -dimensional measurement space involves using an iterative technique known as the Kohonen neural network [41, 42] or self-organizing map (SOM). A Kohonen neural network consists of a layer of neurons arranged in a two-dimensional grid or... 

Jezierska et al. [67] applied the Kohonen neural network to select the most relevant descriptors. Here a Kohonen network is built with the transposed matrix, i.e., with the matrix where the roles of descriptors and molecules are exchanged. From the map of descriptors, 36 descriptors were selected. This number of descriptors was further reduced to six, five, four, or three descriptors. Statistical parameters of compared models are reported in Table 3 of [67]. It is evident from the table that the model built with four selected descriptors show comparable parameters to the model built with 36 descriptors. The selected descriptors belong to topostructural and topochemical classes. 

The Kohonen neural network has been used to develop bioisosteres of the methylen-dioxyphenyl group found in a variety of antag-... 

Cluster analysis methods, principal component analysis and related techniques, and different -> artificial neural networks (such as Kohonen neural networks) are usually used to search for clusters of similar compounds, where a cluster is constituted by distinct objects that are more similar to each other than to objects outside the group. 

Kireev, D.B., Chretien, J.R., Bernard, P. and Ros, F. (1998). Application of Kohonen Neural Networks in Classification of Biologically Active Compounds. SAR QSAR Environ.Res., 8, 93-107. 

A type of neural network that has been proved to be successful in a series of applications is based on self-organizing maps (SOMs) or Kohonen neural networks [61]. Whereas most of the networks are designed for supervised learning tasks (i.e., the relationship between input and output must be known in form of a mathematical model), Kohonen neural networks are designed primarily for unsupervised learning where no prior knowledge about this relationship is necessary [62,63]. 

The Kohonen neural network is similar to a matrix of vectors, each of which represents a neuron. If we arrange the neurons in a matrix and look from the top, we can actually think of a map of neurons where the connectors that have been adapted in the previous model are fixed and depend only on the topological distance of the neurons (Figure 4.12). What are adapted in this model are the components of the vectors in the neural network, which are in this case called weights. Each input is also a... 

Counterpropagation (CPG) Neural Networks are a type of ANN consisting of multiple layers (i.e., input, output, map) in which the hidden layer is a Kohonen neural network. This model eliminates the need for back-propagation, thereby reducing training time. 

Kohonen Neural Networks or self-organizing maps (SOMs) are a type of ANN designed for unsupervised learning where no prior knowledge about this relationship is necessary. 

Training a Kohonen neural network with a molecular descriptor and a spectrum vector models the rather complex relationship between a molecule and an infrared spectrum. This relationship is stored in the Kohonen network by assigning the weights through a competitive learning technique from a suitable training set of... 

We have seen that RDF descriptors are one-dimensional representations of the 3D structure of a molecule. A classification of molecular structures containing characteristic structural features shows how the descriptor preserves effectively the 3D structure information. For this experiment, Cartesian RDF descriptors were calculated for a mixed data set of 100 benzene derivatives and 100 cyclohexane derivatives. Each compound was assigned to one of these classes, and a Kohonen neural network was trained with these data. The task for the Kohonen network was to classify the compounds according to their Cartesian RDF descriptors. 

The reduction of the descriptor size (i.e., the decrease in resolution) usually has a profound influence on the ability of the descriptor to characterize a molecule. Even though compressed, or filtered, wavelet transforms of descriptors have a reduced size, they preserve the similarity information well and in a much more efficient way. Figure 6.15 shows results from an experiment where a Kohonen neural network classifies the same data set (100 benzene derivatives plus 100 monocyclic cyclohexane derivatives) according to ring type. 

Kohonen neural networks are able to map the surface of a molecule together with its electrostatic potential distribution onto a 2D plane the Kohonen map. These maps can be compared since their size is independent of the size of the molecule s surface. The lower right images in Figure 6.38 show an example a pattern of the electrostatic potential mapped into a 2D plane as performed by a Kohonen neural network. This visual representation can be used as a descriptor in a search for compounds with similar biological activity. 

Hasegawa, K., Matsuoka, S., Arakawa, M. and Funatsu, K. (2002) New molecular surface-based 3D-QSAR method using Kohonen neural network and 3-way PLS. Computers Chem., 26, 583-589. 

Mazzatorta, P., VraHco, M., Jezierska, A. and Benfenati, E. (2003b) Modeling toxicity by using supervised Kohonen neural networks./. Chem. Inf Comput. Sci., 43, 485—492. 

---