My next Tweet could be generated by a deep-learning neural network.
2020-07-25 12:07:03
My next tweet could be generated by a deep-learning neural network. I've been training one. Would anyone notice the difference? Could I just hand over tweeting to my machine?
Method:
- downloaded the last 3200 Tweets that I posted using allmytweets.net
- pruned the dates off and removed the RTs by using some regular expressions on a text file
- uploaded the file to Google's Colaboratory and used a notebook template for Max Wolf's textgenrnn as explained in his video
- fiddled with the neural network settings and generated a model of my personal style of Twitter writing (see picture of mid-fiddle)
- downloaded the model to my PC (sadly without a fancy expensive GPU)
- installed Python3, tensorflow, and all the other required libraries
- I am now fiddling with the output. The creativity can be adjusted, as well as the length of the output, and the seed text to get it going
Results:
- the machine can learn at the level of characters (not words) so it is able to put creative new language together.
- it makes new words like 'Wisterage' (presumably from wisteria), 'Cardify' (to be more like something from Cardiff perhaps), and 'crimpumery' (something to do with Cornish pasties?)
- it throws in lots of @ mentions of the people I talk to - possibly because there are multiple occurrences of the same text sequence
- It creates lots of URLs (the shortened ones 't.co' that Twitter stores) but most of them don't work - probably because they are all unique and the text output is creative
Conclusion:
- Early mentions of @dr_mattmorgan and @DrLindaDykes in the output so far
- The curious phrase that I could have written but didn't. "Consultation through artificial learners designers and events. :-)"
Source: agnate.co.uk
output text generated neural network downloaded mentions deep-learning
Related: