Re: Speaker suggestions?

On Fri, Jun 29, 2018 at 07:08:24PM +0000, Karen Shaeffer wrote:
> On Thu, Jun 28, 2018 at 09:01:03PM -0700, Rick Moen wrote:
> > Quoting Bob Smith ([email protected]):
> > 
> > > Let's be generous and say I fully understand
> > > public key encryption.  Starting with that
> > > knowledge, is there anyone on the list who
> > > could explain blockchains?
> > > Not cryptocurrency, just blockchain.
> > > Is there a way do demo blockchain using the
> > > people who attend to validate the transaction?
> > 
> > Bob, is this a speaker suggestion you are making?  Otherwise, IMO you
> > really should start a new thread, to avoid interfering with SVLUG trying
> > to quickly line up a July 2018 speaker.
> > 
> > If you are indeed offering to be that speaker, thank you indeed.
> 
> Hi folks,
> I'm not volunteering to be a speaker. I am encouraging this apparent
> suggestion from Bob that the group get involved in some type of
> blockchain demo.
> 
> My understanding is blockchain is one of the most disruptive technology
> concepts out in the wild today. Some folks have opined blockchain has
> the capacity to bring down the Google and Facebook empires at some
> future date:
> 
> http://rbharath.github.io/why-blockchain-could-one-day-topple-google/

It is instructive to follow this link from that blog:

https://ai.googleblog.com/2017/04/federated-learning-collaborative.html

Here, we see Google's definition of Federated learning. Notice how the
individual perturbations to the AI model at each Android device are
collected at Google's cloud.

This supposedly provides each user some privacy about the details of
their data. But Google has all the differential perturbations. And they
know which device each perturbation came from. And is likely they can
develop AI pattern recognition models to decipher the information that
each perturbation from an individual user translates to in the field of
the user's private data. Afterall, the problem domain is known and has
bounding constraints. And this notion of distributed Federated learning
is going to produce billions of perturbations, each associated with a
known device. And so this is a half step intended to preserve Google's
empire in the face of the long term threat Blockchain represents.

Bharath Ramsundar is saying the model can be encapulated in the
blockchain. And thus the owners of the data who are willing participants
can control who has access to every aspect of their data, including the
model perturbations. Such a framework wouldn't rule out computations on
Google's cloud, but it wouldn't be required. Privacy can be realized
with fine grained control of each aspect of the model, data, training
computations, and inference computations. Will Google's Federated
learning framework win mindshare over pure blockchain solutions in the
future. Who knows. I'm not antiGoogle. Just discussing the issues.

humbly,
Karen
-- 
Karen Shaeffer                 The subconscious mind is driven by your deeply
Neuralscape Services           held beliefs -- not your deeply held desires.