blending authors to make idiot soup with neural networks
I like building deep learning architectures,
to explore what it means to be an algorithm,
and define a bot by what it has experienced.
If we can teach machines to talk like us,
we can use them to create a new medium,
that remixes the past in a new way.
Building an author bot
-
Download a large corpus of text
-
Train a two-layer LSTM RNN over random sentences
-
Use Tensorflow (with Keras) to quickly prototype
-
Fix the top layer and train each author over the bottom layer
now bots share a common vocabulary!
Lovecraft bot
Hemingway bot
Blending bots
-
Sample a character from each bot
-
Choose which bot gets to speak with probability:
-
-
Restart word if it doesn't exist in author vocabulary
-
Parallel sample and choose best paragraph
Shakespeare
to Lovecraft
to Lovecraft