I have been unable to find the article since—I think it must have been Scientific American. Perhaps in the 1980s.

In any event, it described training a neural network, perhaps it was number recognition. The author said that when they "destroyed" the network it began to have "flashbacks" that resembled early training sessions.

That always stuck with me.

▲

zknowledge 2 days ago | parent [-]

Was it this Geoffrey Hinton paper? It was in Scientific American in 1993.

Simulating Brain Damage:

https://www.cs.toronto.edu/~fritz/absps/sciam93.pdf

	▲	JKCalhoun 2 days ago \| parent [-]
		I don't think so. When I just asked an LLM though, the Hinton paper also came up. So it does look like a strong candidate. If I am able to find it, I'll comment here.