The Extraordinary Link Between Deep Neural Networks and the Nature of the Universe

“The an­swer is that the uni­verse is gov­erned by a tiny sub­set of all pos­si­ble func­tions. In other words, when the laws of physics are writ­ten down math­e­mat­i­cally, they can all be de­scribed by func­tions that have a re­mark­able set of sim­ple prop­er­ties.”

“For rea­sons that are still not fully un­der­stood, our uni­verse can be ac­cu­rately de­scribed by polyno­mial Hamil­to­ni­ans of low or­der.” Th­ese prop­er­ties mean that neu­ral net­works do not need to ap­prox­i­mate an in­fini­tude of pos­si­ble math­e­mat­i­cal func­tions but only a tiny sub­set of the sim­plest ones.”

In­ter­est­ing ar­ti­cle, and just div­ing into the pa­per now, but it looks like this is a big boost to the simu­la­tion ar­gu­ment. If the uni­verse is built like a game en­g­ine, with stacked sets like Man­delbrots, then the sim­plic­ity it­self be­comes a driver in a fabri­cated re­al­ity.


Why does deep and cheap learn­ing work so well?