[FRIAM] How does that style transfer work?

[Nick Thompson]

“Hebbian” accuracy?????

 

How does that differ from “Peircean truth”?

 

n

 

Nicholas S. Thompson

Emeritus Professor of Psychology and Biology

Clark University

 <http://home.earthlink.net/~nickthompson/naturaldesigns/> http://home.earthlink.net/~nickthompson/naturaldesigns/

 

From: Friam [mailto:friam-bounces at redfish.com] On Behalf Of Jon Zingale
Sent: Wednesday, October 23, 2019 10:28 AM
To: friam at redfish.com
Subject: Re: [FRIAM] How does that style transfer work?

 

Cody,

 

I really have no idea. The site that I signed up for is called

Deep Dream Generator <https://deepdreamgenerator.com> , and it supposedly exists to give

users hands on experience with image generation with

neural nets. It is setup like the typical social media site,

allowing users to make their `dreams` public and for

others to `like` the `dreams`. Experimenting with different

images and styles is starting to give me some heuristics

to work from and some ideas for how features are extracted.

 

My current working theory is something of a metaphor.

I imagine that when a `style` image is uploaded the

neural net is like a teenager in a punk band. It has maybe

three favorite bands, but one that really really is the best.

The teenagers memorize and test themselves on all the

idiosyncrasies of this favorite band, deconstructing the

lyrics and the moves on stage. Then when teenager and

band of friends go to cover a song by anyone else, this

song is effectively reconstituted from all of this training

on the bands favorite band.

 

Meanwhile, what is actually happening? Well, you know,

weights on edges in a database are increasing when nodes

`fire` together, and weaken when Hebbian accuracy is not

achieved.

 

Jon

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://redfish.com/pipermail/friam_redfish.com/attachments/20191023/9e574da3/attachment.html>