Building Safe A.I. - i am trask


19 bookmarks. First posted by pavelr 6 days ago.


The goal is simple. We want to build A.I. technology that can become incredibly smart (smart enough to cure cancer, end world hunger, etc.), but whose intelligence is controlled by a human with a key, such that the application of intelligence is limited. Unlimited learning is great, but unlimited application of that knowledge is potentially dangerous.

To introduce this idea, I'll quickly describe two very exciting fields of research: Deep Learning and Homomorphic Encryption.
AI  robotics  intelligence  deep_learning  homomorphic_encryption 
3 days ago by naijeru
A machine learning craftsmanship blog.
5 days ago by martinbalfanz
<< With the approach above, you could train a regular, decrypted neural network for a while, encrypt it, send it to Party A with a public key (who trains it for a while on their own data... which remains in their possession). Then, you could get the network back, decrypt it, re-encrypt it with a different key and send it to Party B who does some training on their data. Since the network itself is what's enrypted, you get total control over the intelligence that you're capturing along the way. Party A and Party B would have no way of knowing that they each received the same network, and this all happens without them ever being able to see or use the network on their own data. You, the company, retain control over the IP in the neural network, and each user retains control over their own data. >>

I suspect that this isn't actually going to work but bookmarking for future digestion.
artificial-intelligence  neural-networks  cryptography 
5 days ago by absfac
Building Safe AI: A Tutorial on Homomorphically Encrypted Deep Learning: (
from twitter
5 days ago by electricant
TLDR: In this blogpost, we're going to train a neural network that is fully encrypted during training (trained on unencrypted data). The result will be a neural network with two beneficial properties.
IFTTT  Pocket 
5 days ago by timothyarnold
Building Safe AI: A Tutorial on Homomorphically Encrypted Deep Learning ()
from twitter_favs
6 days ago by ngpestelos
Building Safe AI: A Tutorial on Homomorphically Encrypted Deep Learning.
deep_learning  homomorphic_encryption 
6 days ago by Sylphe