Keynote: computational stichography! (or: connecting lines)
summary
A Gutenberg poetry corpus collects about three million lines (~52 MB) and feeds phonetic vectors (CMU Pronouncing Dictionary, feature counts, PCA to 50D) into a variational autoencoder that generates lines and circular interpolations; a bi-directional RNN trained as a stickography line-break predictor reaches ~93% validation accuracy after overnight training on a 2013 MacBook Air; an nginx+Redis autocomplete and a live typing interface demonstrate the models working on nonsense, Wikipedia prose and Project Gutenberg text.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 1:27The importance of lines
- 4:02Organizing data
- 8:51Linen and cannabis
- 11:23Graphic and poetic lines
- 16:24tidiness and fleshiness
- 24:33Gutenberg Poetry Corpus
- 25:33Poetic Lines
- 28:44Variational Auto Encoder
- 31:51Circular Interpolation
- 34:11Computational stichography
- 40:05Conclusion
related talks
39mExperimental Creative Writing with the Vectorized Word
03h 22mNonsense verse with Python and machine learning
11mMutating text toward better nonsense?!
21mWords Can Communicate Beyond Words - Poetry in the Age of AI
19mSketchforme: Composing Sketched Scenes from Text Descriptions for Interactive Applications
51mPlayful Text
38mMusical Steganography: Hiding Things in Music
31mDoug Eck, Research Scientist, Google