Very interesting article, Jarebb Ng. Your step-by-step explanation suggests a question that I think goes beyond neural networks:
A neural network can correct its error. But can it understand its mistake?
Backpropagation calculates gradients; the optimizer changes weights. The system may consequently predict better next time. Yet improved prediction does not by itself explain whether the original failure came from the data, the representation, a learned shortcut, missing context, or a change in the environment.
Get José. T Guevara’s stories in your inbox
Join Medium for free to get updates from this writer.
That suggests a distinction I find increasingly important:
Error correction ≠ error understanding.
If that distinction holds, the next architectural challenge for AI may not simply be building models that learn from errors, but systems capable of determining what an error means and what should actually be learned from it.
Excellent work. I would be very interested in your thoughts on that distinction. See this publication: https://medium.com/@josetguevara108/when-ai-can-rewrite-ai-who-governs-the-rewrite-6c4469ffb2bd
