The cognitive process of neurologic localization was modeled with a neural network trained on real-world data error backpropagation learning. The task for the network involved learning an internal representation of the mapping from neurological examination findings to the computed tomography (CT) scan lesion distribution implied by those findings in acute stroke cases. The ability of the trained network to localize neurologic exam findings to CT lesion patterns in cases on which it had not been trained was tested. The network showed definite, but limited, ability to generalize from its training experience. There was a tendency for generalization to improve with increasing numbers of hidden units. The possible reason for the limitations of the performance of this model are discussed, and directions for further work are explored. A brief summary of some recent literature on the generalization problem in neural networks is presented.
|Original language||English (US)|
|Number of pages||5|
|Journal||Proceedings - Annual Symposium on Computer Applications in Medical Care|
|State||Published - Dec 1 1989|
|Event||Proceedings: Thirteenth Annual Symposium on Computer Applications in Medical Care (SCAMC-13) - Washington, DC, USA|
Duration: Nov 5 1989 → Nov 8 1989
ASJC Scopus subject areas