I really enjoyed this talk by Andrew. Here's the outline: 0:00 - Introduction 0:54 - Privacy preserving AI talk overview 1:28 - Key question: Is it possible to answer questions using data we cannot see? 5:56 - Tool 1: remote execution 8:44 - Tool 2: search and example data 11:35 - Tool 3: differential privacy 28:09 - Tool 4: secure multi-party computation 36:37 - Federated learning 39:55 - AI, privacy, and society 46:23 - Open data for science 50:35 - Single-use accountability 54:29 - End-to-end encrypted services 59:51 - Q&A: privacy of the diagnosis 1:02:49 - Q&A: removing bias from data when data is encrypted 1:03:40 - Q&A: regulation of privacy 1:04:27 - Q&A: OpenMined 1:06:16 - Q&A: encryption and nonlinear functions 1:07:53 - Q&A: path to adoption of privacy-preserving technology 1:11:44 - Q&A: recommendation systems
And his confidence is barely reaching 0.25. I know it’s not easy to present, but making your temper be a tool for conveying a technical topic is weird way of going.
In the Q+A section- the first question was referring to the vulnerability of reverse modeling. The 13x encryption actually helps you here, most likely making this brute force attack impossible, but it would be a concern once compute/epsilon accounting becomes more available. In terms of institutions to act as a central accounting registry, the obvious answer is some blockchain rollout.
Something like the tangle could further optimize this process by simultaneously addressing the scaling issues associated with blockchain networks during the industrial adoption of DP.
Appreciate the upload. This lecture series is amazing. Dropping a comment about secure multi-party computation. This sounds like an amazing application for the blockchain technology. Dunno if this is brought up after the introduction of secure multi-party computation. had to pause when he described it.
Ultimately AI will protect human privacy (from other humans anyway) because it will replace every institution that requires such information for legitimate purposes, and have no need to share it with any humans to carry out those purposes.
You seriously believe that? What you just described is a society of Big Brother Knows All. Worse: Only Big Brother Knows All, replaces all institutional functions of society, and does not share knowledge. The rest of society would be just individuals with drastically limited knowledge of others. They couldn't even think about rebellion. I don't think that's what you wanted.
@@reinerwilhelms-tricarico344 Sort of? I dunno how I feel about it, but I'm pretty sure that eventually we'll have AI computer programs INSTEAD of governments. And that is kind of a crapshoot. We need to figure out real AI. And we need to do a good job of it, because right after we do, it will figure out us.
@@zrebbesh Before we even try to "figure out real AI" we need to be clear who it serves and who makes the policies. Currently almost all progress in science, especially in computer science, is biased towards serving a small moneyed elite. At the same time we are driving towards a kind of neofeudalism, controlled by finance capitalists and asset managers behind the scene. A utopia that starts like that is almost certainly dystopian: robo-cops, war machines, and complete surveillance of everything including your vital signals. No Thanks.
I vote for liquid as well. He is not himself confident in what he is saying. He is very nervous, as a deep liar is. He is having hard times conveying his message during the whole lecture. My explanation is: he is trying to present an elephant while he only has a mouse in his hands.
popularizing spyware ? Remember up is down, left is right, and the party will tell you what you need to know. Let's install more spyware so that people will have better privacy.
28:00, the data is safe, but the model is put at risk/do a join/computation across multiple data owners. Can't differential privacy handle these? Do DP has these two Cons?
Requires a longer conversation - but the short answer is validation data. There are also ways to try to stretch your validation data to make the best use of it as well.
A good The validation data system already exist? Because if you take information in many form of discipline i dont understand how the system can validate the accuracy of the information (thanks for your time)
Excellent question. I have the same question. Now trust is shifted to thousands of unidentified nodes that may be controlled by a single owner (hacker). Think cars sharing data with insurance company.
diana amiri I’ll try to explain that briefly...that probably means they didn’t like what was shown in the video. There could be a variety of reasons for that. Hope that helps.
Very good talk. But I dont like the way he continues to raise his hand and ask if there anybody knows that or if there anybody knows this. What a embarrassing way to teach.
I really enjoyed this talk by Andrew. Here's the outline:
0:00 - Introduction
0:54 - Privacy preserving AI talk overview
1:28 - Key question: Is it possible to answer questions using data we cannot see?
5:56 - Tool 1: remote execution
8:44 - Tool 2: search and example data
11:35 - Tool 3: differential privacy
28:09 - Tool 4: secure multi-party computation
36:37 - Federated learning
39:55 - AI, privacy, and society
46:23 - Open data for science
50:35 - Single-use accountability
54:29 - End-to-end encrypted services
59:51 - Q&A: privacy of the diagnosis
1:02:49 - Q&A: removing bias from data when data is encrypted
1:03:40 - Q&A: regulation of privacy
1:04:27 - Q&A: OpenMined
1:06:16 - Q&A: encryption and nonlinear functions
1:07:53 - Q&A: path to adoption of privacy-preserving technology
1:11:44 - Q&A: recommendation systems
He seems so exciting to convey his idea to the students, what a researcher ! Great job!
I was there! One of the best talks I have seen
Got his book and will hopefully start exploring it soon :) Excellent talk Andrew - thumbs up!
Excellent talk by Andrew Trask.
Andrew task is already in 2.x
I put the speed on .75, and it was perfect.
And his confidence is barely reaching 0.25. I know it’s not easy to present, but making your temper be a tool for conveying a technical topic is weird way of going.
jeez, I thought you were recommending 2.x and to my surprise he sounded like an alien.
I was actually checking whether my speed was right.
In the Q+A section- the first question was referring to the vulnerability of reverse modeling. The 13x encryption actually helps you here, most likely making this brute force attack impossible, but it would be a concern once compute/epsilon accounting becomes more available. In terms of institutions to act as a central accounting registry, the obvious answer is some blockchain rollout.
Something like the tangle could further optimize this process by simultaneously addressing the scaling issues associated with blockchain networks during the industrial adoption of DP.
I just like the black and white suit Lex Fridman used, very neat and consistent, deep voice and articulation 👍👍👍👍 what an amazing personality
Amazing talk by atrask! I'd love to implement differential privacy for sure!
Sometimes RUclips really suggests exactly what your looking for. Thank you Andrew! I am so grateful I got to listen to your lecture :)
Greatly enjoying your book Andrew!
As someone that works in healthcare, this is fascinating.
great talk!
Appreciated the sound of disgust when he was talking about the company selling de-anonymized data
Could this have any relevance to Coronavirus contact tracing?
Appreciate the upload. This lecture series is amazing. Dropping a comment about secure multi-party computation. This sounds like an amazing application for the blockchain technology. Dunno if this is brought up after the introduction of secure multi-party computation. had to pause when he described it.
What a fantastic talk, enjoyed every minute of it. He also wants the talk himself, which is quite interesting.
The sewage analogy was brilliant.
Excellent talk! Really enjoyed it.
Thank you Lex!
Ultimately AI will protect human privacy (from other humans anyway) because it will replace every institution that requires such information for legitimate purposes, and have no need to share it with any humans to carry out those purposes.
You seriously believe that? What you just described is a society of Big Brother Knows All. Worse: Only Big Brother Knows All, replaces all institutional functions of society, and does not share knowledge. The rest of society would be just individuals with drastically limited knowledge of others. They couldn't even think about rebellion. I don't think that's what you wanted.
@@reinerwilhelms-tricarico344 Sort of? I dunno how I feel about it, but I'm pretty sure that eventually we'll have AI computer programs INSTEAD of governments. And that is kind of a crapshoot. We need to figure out real AI. And we need to do a good job of it, because right after we do, it will figure out us.
@@zrebbesh Before we even try to "figure out real AI" we need to be clear who it serves and who makes the policies. Currently almost all progress in science, especially in computer science, is biased towards serving a small moneyed elite. At the same time we are driving towards a kind of neofeudalism, controlled by finance capitalists and asset managers behind the scene. A utopia that starts like that is almost certainly dystopian: robo-cops, war machines, and complete surveillance of everything including your vital signals. No Thanks.
30:38 will this work with non-linearities which are quite important for neural nets?
I guess 1:06:38 has the same question, nevermind
The most complex thing here is to not look at the line integral on the blackboard:))
Solid
@@sp1534 nah.. solid really
I vote for liquid as well. He is not himself confident in what he is saying. He is very nervous, as a deep liar is. He is having hard times conveying his message during the whole lecture. My explanation is: he is trying to present an elephant while he only has a mouse in his hands.
popularizing spyware ? Remember up is down, left is right, and the party will tell you what you need to know. Let's install more spyware so that people will have better privacy.
28:00, the data is safe, but the model is put at risk/do a join/computation across multiple data owners.
Can't differential privacy handle these? Do DP has these two Cons?
Maybe is a bad question, what happen if massive malicious group enter hudge quantity of fake data in that kind of system?
Requires a longer conversation - but the short answer is validation data. There are also ways to try to stretch your validation data to make the best use of it as well.
A good The validation data system already exist? Because if you take information in many form of discipline i dont understand how the system can validate the accuracy of the information (thanks for your time)
Excellent question. I have the same question. Now trust is shifted to thousands of unidentified nodes that may be controlled by a single owner (hacker). Think cars sharing data with insurance company.
I don't understand people that disliked!!!
diana amiri I’ll try to explain that briefly...that probably means they didn’t like what was shown in the video. There could be a variety of reasons for that. Hope that helps.
I can give you a few reasons
Singularitynet are building a Decentralised AI Ecosystem Built on these ideas. A Must to check out!!!
Great Talk. Regarding sewage system and water supply, have a look what the Romans were capable of doing. They were 1000s years ahead of their time.
I love this
heeey im getting 4 mid video ads in one go now ;( (great intro to ppML)
Cool to see Nick Mullen has gotten into computer science
He is talking so fast. Or is the video accelerated by default?
video went 2x after Lex introduction.
"Trask" eh? Builds robots, eh? No way this ends poorly.
4:06 The common cold... Kind of far sighted
"Does that make sense?"
...um
Very good talk. But I dont like the way he continues to raise his hand and ask if there anybody knows that or if there anybody knows this. What a embarrassing way to teach.
Hey! Thanks for the feedback and I'll reconsider this approach in the future. :)
Machinegun....
This guy looks like peter petrelli
Bennetians coming here during endterms
He spoke to quickly, don't know what was pursuing him. He could have been signaled to slow down a bit.
The original playback is fast, you can re-set playback speed to .75 (bottom right hand corner of screen) it will be easier for you.
He seems so exciting to convey his idea to the students, what a researcher ! Great job!