#wkto

[baconmancr]

[baconmancr]

Spoiling the punchline: trained neural networks tend to generalize because good, generalizing minima have basins of attraction that are 80,000 orders of magnitude larger than the bad minima.

For reference, WolframAlpha tells me that the observable universe is about 184 orders of magnitude larger than the Planck volume.

[baconmancr]

[baconmancr]

[baconmancr]