Work began on this project during FY92. Its goals are: (1) to provide services to members of the NIH research community exploiting artificial neural network (ANN) computational paradigms. These applications range from noise elimination to medical decision making to evaluation of screening of therapeutic agents. ANN methods are attractive alternatives in some problems also approachable by traditional statistical methods, especially when the need for cheap rapid aquisition of a model outweighs the need or precise measurement of error or confidence. The project will study and support, for the NIH community, the most useful generalized ANN software platforms, consult with, and may collaboratively assist biomedical scientists to utilize ANNs. A bi-weekly network interest group in a journal club format begain with a nucleus of interested NIH colleagues. A 4-day neural network course, taught by an outside expert, and open to the NIH community, began an educational effort in September. (2) Studies by LSM staff included surveys of existing ANN applications in two fields: medical diagnosis and DNA protein sequence analysis. Preliminary formulation was done for algorithms for training neural networks with unequal error weighting (e.g., when screening for a serious illness the error of falsely diagnosing the illness might be judged preferable to the error of failing to detect it), and for the learning of """"""""hard"""""""" Boolean functions (a bid to raise the level of Kolmogorof complexity in data which neural net models are able to represent).
