“Harmony” was also the metric used for the first multi-layer neural network by Smolensky (1983). Then Hinton called it (free) “energy”—same metric with a sign r

“Harmony” was also the metric used for the first multi-layer neural network by Smolensky (1983). Then Hinton called it (free) “energy”—same metric with a sign reversal. Then, with Rummelhart, they compromised and called it “goodness of fit.” Same as Harmony. So, neural nets still optimize for Harmony, by another name. I find that wonderful, somehow.