Digital art based on Deep learning

http://ml4a.github.io/classes/itp-F18/

https://nips2017creativity.github.io/

https://janhuenermann.com/blog/abstract-art-with-ml

https://janhuenermann.com/blog/abstract-art-with-ml

Machine-learning extrapolation of art: http://extrapolated-art.com/

ASCII art with deep CNN https://github.com/OsciiArt/DeepAA

https://deepart.io/

Random Pics Combined Using Neural Network

Neural doodle

Colorize B&W pictures: http://demos.algorithmia.com/colorize-photos/

Make photo from segmentation: http://prostheticknowledge.tumblr.com/post/169038480796/uncanny-rd-project-by-anastasis-germanidis-and

Generative adversarial networks

big gans https://mobile.twitter.com/neuroecology/status/1073291777321381888

make.girls.moe / Crypko

https://dena.com/intl/anime-generation/

Neural Style: A Neural Algorithm of Artistic Style

https://blog.openai.com/glow/

Deep dream

Inceptionism: Going Deeper into Neural Networks – Going Deeper with Convolutions

How ANYONE can create Deep Style images

https://twitter.com/quasimondo https://twitter.com/mtyka https://twitter.com/alexjc https://twitter.com/Salavon https://twitter.com/genekogan https://twitter.com/chrisrodley https://twitter.com/elluba https://twitter.com/dh7net https://twitter.com/kostiumas

https://twitter.com/artwithMI https://twitter.com/ml4a_

https://twitter.com/hardmaru https://twitter.com/samim https://twitter.com/algoritmic https://twitter.com/prostheticknowl https://twitter.com/fchollet https://twitter.com/zachlieberman https://twitter.com/bitcraftlab

Style Transfer for Headshot Portraits

Deep Convolutional Inverse Graphics Network

Image based relighting using neural networks

pix2pix

https://magenta.tensorflow.org/welcome-to-magenta

Deep learning for Music

https://mobile.twitter.com/chrisdonahuey/status/1073387592161193984 Music transformer https://magenta.tensorflow.org/music-transformer

https://www.youtube.com/watch?v=HANeLG0l2GA

Composing Music With Recurrent Neural Networks – https://affinelayer.com/sidgen/

https://www.technologyreview.com/s/603137/deep-learning-machine-listens-to-bach-then-writes-its-own-music-in-the-same-style/

Music recommendation

Long tail (Power laws) is particularly long in music

• Collaborative filtering. Problem with new users. Performs the best.
• Content-based. No popularity or user usage data required.

Diversity is valuable in long term.

Deep content-based music recommendation.

Model with latent factors. Model from raw audio signal to latent space. Map users to latent space, and see songs nearby, to recommend. Use ConvNet to implement the map from audio signal to latent space.

Weighted Matrix factorization (uses confidence matrix) to find the latent factors (see [[Compressed sensing] video]). Basically implements: probability of having listened to song by vector product between user and song latent representations.

Mean squared error is rotationally invariant, and factorizations are too!

Latent factors split into those predictable from audio, and predictable from metadata.

Datasets: million song dataset, echonest.

Still much poorer performance than collaborative filtering, in general.

Can visualize, using T distributed stochastic neighbour embeding, can identify some genres.