Neural networks: New tools for modelling and data analysis in science

To provide a primer for the study of scientific applications of connectionist systems, the dynamical, statistical, and computational properties of the most prominent artificial neural-network models are reviewed. The basic ingredients of neural modeling are introduced, including architecture, neuronal response, dynamical equations, coding schemes, and learning rules. Perceptron systems and recurrent attractor networks are highlighted. Applications of recurrent nets as content-addressable memories and for the solution of combinatorial optimization problems are described. The backpropagation algorithm for supervised training of multilayer perceptrons is developed, and the utility of these systems in classification and function approximation tasks is discussed. Some instructive scientific applications in astronomy, physical chemistry, nuclear physics, protein structure, and experimental high-energy physics are examined in detail. A special effort is made to illuminate the nature of neural-network models as automated devices that learn the statistics of their data environment and perform statistical inference at a level that may approach the Bayesian ideal. The review closes with a critical assessment of the strengths and weaknesses of neural networks as aids to modeling and data analysis in science.