How to evaluate a neural network without training?

Question

Neural networks

How to determine the productivity/efficiency of a neural network by its structure? I can drive literally a couple of training epochs and get a dynamic error, but the network has not been brought to the "mind", how can I get its estimate here? I thought something like the learning rate, but after all, one neuron will quickly approach the goal for the first 10 minutes, and then it will slip through, and the other will be slower, but will come right to the target, the learning rate standard.

How to determine the efficiency of a neuron by its architecture (the number of layers and neurons in them) using almost no training or using, but not perfecting the grid.

Answer 1

Or how can I understand, having only the structure of a neural network, that it can even solve my problem? Type linear data is enough and 0 hidden, and if far from linear, how to understand.

Answer 2

No way

Answer 3

Just like determining a person's QI by a smile.
It is worth looking through at least one work about neural networks, and you will see comparative tables there, where they played with some parameters, others, network structure, input data, etc. - all this affects the result. And here you want to be given advice in two sentences.