News for hackers. Stuff that matters.

A comment on Building Safe AI: A Tutorial on Homomorphically Encrypted Deep Learning.

Re:Building Safe AI: A Tutorial on Homomorphically Encrypted... by javajosh on Friday March 17, 2017 @08:10PM (#13899091)
Seems to me that dropping the terms of a Taylor expansion could have wide-ranging consequences to the coherence of an artificial mind, making this approach infeasible.
[ Parent | on HN ]
Re:Building Safe AI: A Tutorial on Homomorphically Encrypted... by jordancampbell on Saturday March 18, 2017 @03:03AM (#13900258)
In general you don't actually need crazy precision to train the nets, and a small number of Taylor expansion terms tends to approximate functions fairly well anyway.
[ Parent | on HN ]