05-12-2025, 11:14 AM
Hi all. 
Finally I have published a new version of my program "NeuroLab". DWXL365.000-2
(short "User's Guide" - DWXL365.000-1 )
This version allows you to conduct some experiments and draw some conclusions.
It is already clear to me that a neural network capable of recognizing figures by features must have neurons in its first hidden layer that are trained to recognize not only the presence of features in an image, but also the location of the features and the distance between the features.
Only then can the second hidden layer of neurons be trained to detect the shape in the image using the data from the first layer.
At this point, I still hold the opinion that layers of neurons in a multilayer network should be trained one by one, rather than all layers at once. Otherwise, we train neurons, for example, of the second layer ON UNRELIABLE DATA, which the second layer receives from the still untrained neurons of the first layer.
Is this logical?..
Finally I have published a new version of my program "NeuroLab". DWXL365.000-2
(short "User's Guide" - DWXL365.000-1 )
This version allows you to conduct some experiments and draw some conclusions.
It is already clear to me that a neural network capable of recognizing figures by features must have neurons in its first hidden layer that are trained to recognize not only the presence of features in an image, but also the location of the features and the distance between the features.
Only then can the second hidden layer of neurons be trained to detect the shape in the image using the data from the first layer.
At this point, I still hold the opinion that layers of neurons in a multilayer network should be trained one by one, rather than all layers at once. Otherwise, we train neurons, for example, of the second layer ON UNRELIABLE DATA, which the second layer receives from the still untrained neurons of the first layer.
Is this logical?..