Paper: Multiple disorder diagnosis with adaptive competitive neural networks

摘要: Backpropagation neural networks have repeatedly been used for diagnostic problem-solving, but not demonstrated to work well when multiple disorders are present. We hypothesized that letting nodes in a backpropagation network compete be part of solution would produce better performance than the use existing methods. To test this hypothesis, we derived an error learning rule can with competitive units (competitive backpropagation). Artificial were then trained using both new and standard on specific medical diagnosis problem: identification location damage brain given set examination findings. Training samples included solely 'prototypical' cases where single is The tested atypical manifestations more one disorder present or only manifestation was Networks employing competition among found perform qualitatively these multiple-disorder also single-manifestation cases. reasons explained. described here provides promising tool adaptive problem-solving.