3 ms·

I think it has more to do with the content itself, and the elasticity and adaptability of neural networks. There is no "up" or "down" with inputs. I learned thi

I think it has more to do with the content itself, and the elasticity and adaptability of neural networks. There is no "up" or "down" with inputs. I learned this when working with a machine vision problem. A model will adapt to the orientation or format of a given input. The orientation itself is never a variable, the network simply doesn't care and adapts to it. If you train a model to recognize input in a given orientation, any other orientation will give a high error rate. If you train the model with _every_ orientation, it will be able to classify inputs that are upside down, sideways, mirrored, etc., all relative to the original orientation.

foobarbecue 4y ago
Exactly. There is no "upside-down" unless you train one way and then flip it. So I wish people would stop saying "babies see upside-down at first."
- a9h74j 4y ago
  To your point ... FWIW, young kids in cultures where they were carried in all different orientations as infants allegedly would not turn over a picture handed to them "upside down" to look at it - whereas kids who were held more upright as infants would. Presumably the former could make sense of objects in many more orientations. (Somewhat sure this was in Ashely Montegue's treatise on Touch.)