What if Picasso was a robot? Teaching art to the machine
summary
Trains a deep convolutional GAN to generate 64×64 artworks, finds Picasso yields ~1k images—too few—so crawls 220k landscapes from WikiArt, preprocesses into .npz, and trains with Chainer/CUDA; training runs ~16 hours on 4GB GPU. Results produce mixed-quality images—some convincing, many blurry—and certain genres (mythology) fail due to style variance. Model converts via Keras/TensorFlow to run as a TensorFlow.js browser demo, showing generation works but remains unpredictable.
AI-written summary · about
chapters (20)
- 0:00Start
- 2:26Teaching the computer how to make art
- 4:15One and Three Chairs - Joseph Kosuth 1965
- 4:33Kosuth didn't make the chair, take the photo, or wrote the definition
- 4:52What is machine learning?
- 6:07Several kinds of algorithms to each kind of problem
- 6:50Deep Convolutional Generative Adversarial Networks
- 7:01Capable of creating new images from existing ones
- 7:32convolutional neural networks that keep battling each other
- 7:46Baby steps!
- 10:22Round 2
- 12:4024.832 images of landscapes
- 14:34GB GPU
- 16:08Stopped to make some adjustments, repeat cycle
- 19:42High variance of quality in each generated art work
- 21:00We have to get used with a future where machines can also do creative work
- 23:02There's always the training phase, but also a certain creative unpredictability
- 24:17Art is making meaning
- 24:31What's art to you?
- 25:03THE BIGGEST NODE.JS CONFERENCE
related talks
57mWhat if Picasso was a robot? Teaching art to the machine
56mPicasso's terminal; data science and AI in the visual arts
31mPicasso, Geometry, Jupyter
36mArt in the Age of Algorithms
06mPainting Mona Lisa with Artificial Intelligence
40mLearn to be a painter using Neural Style Painting
39mKeynote: Artistic Applications of Artificial Intelligence
32mCan AI Be Creative?