What if Picasso was a robot? Teaching art to the machine
summary
Uses deep convolutional generative adversarial networks on ~20k scraped WikiArt images (shrunk to 64x64), trains for ~16 hours on a GPU with CUDA, converts a Chainer/Python model into TensorFlow then to TensorFlow.js in the browser. Code pipeline is Node.js scraper → Chainer trainer → model conversion → browser JSON runtime. Output shows 80–84 generated landscape samples with high variance: some convincing, some blurry; success depends on dataset homogeneity, GPU time and layer/weight tuning.
AI-written summary · about
chapters (14)
- 0:00Start
- 4:44One and Three Chairs - Joseph Kosuth 1965
- 5:37What is machine learning?
- 6:59We can use machine learning to draw conclusions about large chunks of data automatically.
- 8:49Our mission
- 9:05Deep Convolutional Generative Adversarial Networks
- 9:27Capable of creating new images from existing ones
- 10:26convolutional neural networks that keep battling each other
- 10:40Each algorithm works based on a model
- 28:42Stopped to make some adjustments, repeat cycle
- 34:17High variance of quality in each generated art work
- 36:17Creative Computers
- 39:03There's always the training phase, but also a certain creative unpredictability
- 41:17Thank you very much!
related talks
25mWhat if Picasso was a robot? Teaching art to the machine
56mPicasso's terminal; data science and AI in the visual arts
31mPicasso, Geometry, Jupyter
25mDecoding Algorithmic Artistry: Visual Generative AI
01h 03mIt's All About AI
36mArt in the Age of Algorithms
38mMaking Art with Artificial Intelligence: Artists in Conversation
58mStyle Transfer AI - How Neural Networks Generate Art