Deep Writing Planet Earth

I think artificial intelligence is really cool because I believe machine learning is transforming the way we use data. I also think that machine learning can be incredible entertaining, especially when living alone in a tiny studio during a pandemic.

For those who don't know, a wonderful world of useless AI applications exists. I'm talking about using deep learning to write a Seinfeld episode, to name cats, and to write the next Harry Potter book (we don't need JK Rowling anymore anyway). But how can we apply this to nature conservation? Luckily for you, I figured out a way...

I present to you Deep Planet Earth:

Follow Fake Nature on Twitter to see more.

---

For those interested, here is the final code. For those even more interested, here is how I wrote it:

I began by following this tutorial and used this guide to set up my environment.

I created my corpus from every episode of Planet Earth I (found here) and II (found here). I copy/pasted the text into a txt file and followed this question/answer to upload the txt file to my Jupyter notebook.

I then followed this TensorFlow tutorial to generate the new Planet Earth episode.

I then experimented with different starting words, epochs and temperatures. Refining the algorithm to avoid overfitting but still generating something that wasn't complete nonsense took some time. 40 epochs, 0.6 temperature, and a starting word of "Here" finally produced excellent results.

With a masterpiece in front of me, I decided to reach out to Reddit to find an Attenborough impersonator who would read my silly script for free. I struck gold. I then turned to royalty free nature videos on Pexels and Pixabay and used free, online video editing software (Unscreen and Kapwing) to bring the script to life. BenSounds provided the royalty free backing music.

After creating Deep Planet Earth, I decided to expand the corpus to include animal facts (source 1 and 2) and 13 PBS nature documentary scripts (source 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13). The "facts" generated are now broader, including a wider array of animals and locations. I rerun the code every so often and list the funniest statements in a Google Sheet Twitter bot that auto tweets to the @_FakeNature account every 6 hours.

What's in the future for Fake Nature? A web app is in the works for users to generate "facts." The user will be able to rate the "facts" as good or nonsense, adding the good statements to the Google Sheet and training the algorithm to be better. This project was intended to be a fun learning tool, but I welcome collaborators to help implement the final stage of the project.