Inside DNN "black box" - recognizing an object - towards a "white box" approach, part 2
Neural networks, especially Deep Neural Networks can be treated, using the terminology of cybernetics - as black boxes. But not necessarily always.
Krzysztof Michalik
12/18/20243 min czytać
In addition to the photos from the 1st part, I have prepared photos from my neural network (DNN) with a sampling period of not 1 epoch but 10 epochs, with the same network topology. At the same time, I have changed the colors of the presentation to shades of gray - colors are important for better perception of some patterns and features created during network training. Btw, in this way you can also debunk many nonsenses usually said by scientists, but not from the field of computer science and AI, but e.g. philosophy or especially physics (they speak on every topic). In one of such statements, the physicist stated that the current AI has reached a level when we no longer understand how it works (sic!). Really? Does this physicist think that AI (like various fairy tales from the field of quantum physics) is already magic and an occupation for Harry Potter, or is it still science?! Similarly, we are being told that a cat can be alive and dead at the same time - Schroedinger himself invented this paradox because he saw the nonsense of this situation, but some quantum mechanics probably took it literally. So maybe something like my modest system for teaching object recognition with the ability to "look inside the network" will protect my students, as well as novice scientists from such a magical narrative? Maybe building such systems, apart from the scientific aspects of their other use, has a deep didactic meaning in disenchanting "AI whose operation we do not understand"?
The photos below were sampled on 3 dense layers every 10 epochs, where epoch 50 practically trained the classifier sufficiently. And it's also time for a photo of one of the heroes of this adventure, the Cat - Sample. :)
Our friend Sample the Kitty
