Potiha study neural networks. The principle I understood, the basic concepts seem to be caught. Even the simplest prototypist wrote.
But what question did I have:
I see that my neural network is not smart enough, no matter how trained, and I understand that it is necessary to increase the number of neurons. But how can I increase their number? Increase the number on the layer, or increase the number of layers?
(Understandably, this needs to be selected experimentally , but due to lack of experience I cannot refer to my observations as valid, and the question is tormenting now).
If I understand correctly, you can imagine an analogy with image processing:
a) Suppose we initially worked only with a binary image view. Well, white / black pixel. And then we want to process the usual black and white image (or color). The number of input neurons may not increase, but we understand that one step has become more complicated, and therefore we can add additional neurons to the existing layer .
b) We want to highlight sharp borders in the image. Understandably, there will already be much more steps , and not just one volume. Here you first need to make black and white from a color image, then, for example, take a square of color, or something similar, in a word, the task is not one step. And here we are already increasing the number of layers in the network.
Is my train of thought correct? Maybe there are some more specific information on this topic that I missed / did not find?
PS for it became the task of making AI for a simple logical game, and I wanted to try to make the computer learn by playing against the computer. And since the process is not fast, I try to estimate the required number of network layers for the AI of this game. I myself meanwhile regard the complexity in a couple of hidden layers, relying only on intuition and an approximate train of thought of the person playing it.
