rga somatic hackday

Post on 15-Apr-2017

223 Views

Category:

Technology

0 Downloads

Preview:

Click to see full reader

TRANSCRIPT

How Neural Networks Can Help Us Expand Visual Creativity

Jason Toy

ABOUT ME

by Jason Toy ❖ builder of things

❖ SFPC

❖ somatic

ArtTech

DEEP DREAMING - INCEPTION

seeing what the computer sees

neural network concepts❖ blackbox

❖ neurons-inputs;outputs; 0-1

❖ accelerates at processing images

❖ layer types

❖ training ~= compiling

❖ most ML problems: classification,regression,reinforcement

deepdream key modifications

how do we understand neural network?

objective function: activate as many neurons in a layer

key trick: push back to image

looping and recursion

GOOGLENET

m

how we changed the objective function of deep dream

❖ https://github.com/somaticio/bat-country/blob/master/batcountry/batcountry.py

STYLE TRANSFER

neural style details

❖ classifier

❖ same network can be used (vgg, googlenet)

❖ visual representation of our world

❖ repurposed - transfer learning

high layers in the network act as the content of the image

style computed from multiple layers’ filter responses

Trials from experimentation of modifying data and algorithms

other examples of image processing networks

❖ black and white photos -> colorized images

❖ low resolution -> super resolution

❖ blurry images -> clear images - CSI magic!!!

❖ image to story -> romance novels / ad generators

Ideas for other transformations?

❖ selective region of inception/style transfer?

❖ dreaming different objects

❖ mixing transformations

❖ really up to our imaginations!! - we will explore

hackathon projects

❖ style transfer photo gallery

❖ slack stylebot

❖ inceptionism - modify an objective function (advanced programming and math)

Thank you!!Questions?

Jason Toyjason@somatic.iohttp://somatic.io

top related