Claude Cachet

Neural network approach to structural feature recognition from infrared spectra

Neural networks, with and without hidden nodes, have been trained to recognize structural features of compounds from their infrared spectra. The training of the networks was evaluated by a variety of statistical indices using threshold values obtained by simplex optimization and by evaluation of synthetic spectra of structural groups obtained from the connection weights of the single-layer networks. Results indicate that all of the networks can be trained to recognize the structural groups in the compounds used to train the network. The network with a hidden layer, and dedicated to a single structural group, was better able to recognize structural groups in compounds that had not been used in training the network. Although not as efficient, the single-layer networks are particularly useful in that information may be extracted for use in writing more effective rules for an expert system-based infrared interpreter.

Advantages of a hierarchical system of neural-networks for the interpretation of infrared spectra in structure determination

Abstract A hierarchical system of small feed forward neural-networks is used to extract structural information from infrared spectra. The top-level network gives a rough classification in five non-exclusive classes: compounds containing carbonyl, hydroxyl, amino groups, aromatic compounds and ethylenic compounds. For each class, a dedicated network is designed to identify more specific structural features. Depending upon those structural features, the hierarchy might be extended to deeper levels. Specialised networks are activated in a cascade-like effect by the output of the upper-level networks. The training of each specialist network is performed using learning and test sets made of compounds identified by the upper level networks as belonging to this class. Thanks to this approach and to the optimisation of decision thresholds, the quality of the responses is excellent, and compounds wrongly classified by one network do not lead automatically to other errors. One major advantage of this approach is the limited size of each network involved. Networks with few outputs are easier to optimise, and their performance is better than that of larger networks. Moreover linking the response sets from the different refinement levels allows improvement of response quality and in some cases inference of other structural features by combination of responses. Hierarchical neural-network systems are well suited for the interpretation of infrared spectra. They perform very well, and the different refinement levels of information permit great flexibility in the ways they may be used. The modular organisation allows modification of some parts of the system without damaging the whole hierarchy.

Indices for the evaluation of neural network performance as classifier: Application to structural elucidation in infrared spectroscopy

An application of Artificial Neural Networks (ANN) to the substructure detection, from infrared spectra, of organic compounds is described. Several ANNs have already been implemented for this purpose, and show promising initial results; however, many problems remain to be resolved. We wished to train ANNs to assist in a decision support system using several spectroscopic methods to elucidate the structure of unknown molecules. To optimise the ANN with respect to spectral feature extraction, network architecture, training regime and threshold determination, we have investigated several indices for use in the evaluation of network performance. Since much published work on ANN application in this field present performance indices that are poorly defined or of limited use, we recommend that the basic results be reported so that readers may calculate indices to suit their own particular needs. These basic quantities are identified and a set of derived indices recommended.

A heuristic problem solver: GEORGE

Abstract Developing problem solving skills is an important educational objective in science teaching. In conventional applications of Computer-Assisted Learning, students are trained through tutorials, drill and practice, guided problems and simulation. One obstacle in the way of achieving a high degree of individualization is that the computer has been unable to answer questions for which answers have not been encoded. The use of artificial intelligence techniques and in particular of expert systems, may remove this obstacle. The program described here constitutes one step in this direction. The program is called GEORGE and has some of the essential characteristics of expert systems. It has been designed to find the solution to most problems of elementary chemistry dealing with mass, volume and number of moles. Contrary to conventional algorithmic programs which deal repeatedly with similar data and always process them in the same way, GEORGE uses heuristic rules to discover a solution to a problem. The program has no questions to offer students but shows users how to solve problems of their own. The heuristic rules are very simple and can be understood by students. In order to use this program, a student must be able to define precisely the quantity which is to be found and to identify the available data. If the available data are insufficient to support a solution, the program tells the user and asks for data relating to the missing information. If the data are sufficient, the program supplies the answer, but more important is that it explains how the answer was reached. Diagrams are used to show the network relating the available items of information to the solution. If the student finds some problems especially interesting. these problems can be saved to be used again or even modified for later use. Thus the student can create a personal collection of problems. This approach is different from traditional categories of Computer-Assisted Learning. The possible impact of creating and using problem solvers in science education is discussed.

Le Système ESSOR d'Expérimentation Scientifique Simulée et ses Potentialités Pédagogiques

Summaries English One of the main aims in the teaching of science is the acquisition of an experimental methodology. The principal obstacles to this methodological training are briefly analysed; they arise mainly from the shortage of time available for experimental work. In this article, computer‐simulated scientific experimentation is proposed as an aid in overcoming these obstacles. This type of teaching, integrated into a classical curriculum, does not exclude laboratory work, which is still indispensable for the learning of techniques. To facilitate the setting up of the proposed method, the ESSOR system simulation on mini‐computer has been developed. Its technical and pedagogic characteristics are described. This system allows the easy simulation of experiments analogous to those of the laboratory. The student using it is free to adopt individual procedures for the study of a proposed phenomenon; the system follows up automatically the individual activity of students and frees the instructor from mos...

Clean-up of Spectro-Structural Databases using Artificial Neural Networks

Quality control in the construction and usage of factual databases is a well known problem. Classical methods of data inspection are not adequate for spectro-structural data of a large number of molecular spectra of various origins (IR, MS, NMR, etc.). Artificial neural networks can be successfully used as non-linectr mapping devices between spectroscopic and structural features. We have built and trained hierarchical neural networks to recognize the presence or absence of several functional groups in a molecule from its infrared spectra Thanks to the speed of computation of these neural networks, it is possible to scan a large spectro-structural database in order to identify doubtful spectra and/or structures. Typically, these represent only two to 15% of the records.