3 ms·
This reminds me a first perceptron tutorial. An author described that to match a digit 9, ideally one layer matches a circle, another sees a lower curve and a r
by wruza 6y ago
This reminds me a first perceptron tutorial. An author described that to match a digit 9, ideally one layer matches a circle, another sees a lower curve and a result is then combined into output “neurons”. In reality, after you train a network, it has no circle-matching or curve-matching parts. Every part of it is just a random noise that you couldn’t reverse-engineer without pushing various data to that blackbox. We could’t even tell if matching F to 1 was a learning error or an intended behavior.