How to properly train a classifier?

Question

Neural networks

Good day. At the moment, I need to create an image classifier that will detect suicidal pictures (cut veins, gallows, blue whales, etc.).
After studying several articles, for example https://habr.com/ru/post/321834/ , a question arose. As I understand it, the classifier refers the picture to one of the classes (in this article, to cats or dogs). And actually it is clear if it is a test on such pictures. And if you give the classifier some kind of picture, like a person / tree / furniture, then what will happen? and how then to train the classifier, where in fact there should be two classes: suicide and everything else?

Answer 1

Yes, to study ML by "articles" is a tin.
"if you give the classifier some kind of picture, like a person / tree / furniture" - do not give it, but first train it on these pictures, then it will classify it into "like a person / tree / furniture". And if you give this picture to a classifier trained on "cats and dogs", then don't be surprised if he takes your stool for a dog, and a tree for a cat. Or vice versa.
how then to train the classifier, where in fact there should be two classes: suicide and everything else - absolutely the same as for cats and dogs, where there are also two classes of them, about which your "articles" write. Let them learn on a couple of tens of thousands of marked up pictures, in which these are your "veins, gallows, blue whales" and the same number of pictures where they are not. So your trained classifier will then divide the pictures into "su e cide and everything else."

Answer 2

any task of building a classifier starts with building a corpus - the initial dataset for analysis - in your case, pictures.
that is, if you want to teach to recognize cut veins on your hands - you need to have 10,000 pictures of cut veins and 10,000 - whole hands,
I think in a specific task there is a very large spread of potential images, very weak signs that need to be identified and therefore nothing will work.