Janelle Shane @JanelleCShane Research Scientist in optics. Plays with neural networks. Avid reader, writer, and player of Irish flute. she/her. wandering.shop/@janellecshane Sep. 15, 2018 1 min read

I fed a neural net 270,000 words of Trump campaign speeches for the @NewYorker and nobody can ever make me do it again.  https://www.youtube.com/watch?v=EFHyzuqjaok 

Methods: Individual sentences (and sometimes groups of sentences) are exactly as the neural net output them. I did select the most interesting sentences, though, and arrange them in an order that (if not exactly making sense) at least flowed better.

Neural nets used:
Writes text letter-by-letter:  https://github.com/karpathy/char-rnn 
Writes text word-by-word:  https://github.com/larspars/word-rnn 
The outputs with nonsense words were mostly from the letter-by-letter neural net.

What the creativity level means: This is the "temperature" setting. At the lowest temperature the neural net always writes the most likely next word/letter, and everything becomes "the the the". At the highest temperature, it chooses less probable words/letters for more weirdness

Training time: like 2 solid days apiece on AWS's Deep Learning AMI.
Sampling time: waaay too long reading Trumplike speech. I have suppressed the memories and it's probably better that way.
Filming time: I dunno; wasn't there, but John Di Domenico absolutely KILLS it

That's in contrast to this neural net (not mine) which trained for over a month on 82 million Amazon reviews.

Via @Johnnyd23 ‘s amazing performance you can see how well the neural net mimics surface characteristics like word choice and rhythm

You can follow @JanelleCShane.


Tip: mention @threader_app on a Twitter thread with the keyword “compile” to get a link to it.

Enjoy Threader? Become member.