The problem is in the implementation of the neural network for pattern recognition, the problem is in the architecture?

Question

Machine learning

Good afternoon, Habr! In neural networks recently and wanted to try to do an interesting but simple task and got stuck in the implementation. The essence of the problem is in the recognition of images of medical packages. I trained the network up to 75-80% (it was even up to 90-99%), but in practice it gives a bad result. I would like someone to consult on this issue. Maybe you should just use a different architecture or fix something in the current one. Thank you for your attention!
Happy coming everyone!

Answer 1

Have you tried reading books? Where the basics of neural networks are neatly and systematically stated, when what architecture to use, what problems, nuances and pitfalls there are, how to improve accuracy and not get into retraining. And how to "fix something" if something went "wrong". All in all, a good textbook. Moreover, such books as rules are written by people who are exactly in the subject and who, at the same time, can also express their thoughts, competently and consistently. And their editors-reviewers - as a rule, they proofread such books and removed obvious errors.
Or reading is not our way. You yourself need to understand, analyze, compare. It’s easier on the forum to ask someone you don’t know, get advice from him of incomprehensible quality and spend time on another iteration with an unpredictable chance of success. And then - either quit this business, or still sit down for books.
And you with the upcoming ones!

Answer 2

Maybe you should just use a different architecture or fix something in the current one.

Might be worth it. Please for your attention!
But seriously, maybe it's worth either concretizing and localizing the problem, or giving the task to a specialist?