Advanced supervised learning in multi-layer perceptrons to the recognition tasks based on correlation indicator

The article deals with the problem of recognition handwritten digits using feedforward neural networks (perceptrons) a correlation indicator. proposed method is based on mathematical model network as an oscillatory system similar to information transmission system. uses theoretical developments authors search for global extremum error function in artificial networks. digit image considered one-dimensional input discrete signal representing combination "perfect writing" and noise, which describes deviation implementation from writing". ideal observer criterion (Kotelnikov criterion), widely used systems probability correct signal, form loss function. In carried out comparative analysis convergence learning experimentally obtained sequences basis indicator tasks classification CrossEntropyLoss use optimizer without it. Based experiments out, it concluded that has advantage 2-3 times.