Down slope doesn't lead to useful model

[I159]

A model recognizes images but this recognition is very low quality - it doesn't reached even 20% after 16 epochs by 64 items per batch and down slope touched it's minimum. Check corrections application correctness.

[I159]

I'm looking for algorithmic issues which could cause low quality of recognition. After refactoring and decomposition I'll implement documentation to a neurons level. Each advice and question is appreciated. Thanks!