This week in our Friday Faves, we have a new Python package for seismic compression and input/output, a conversation on the merits and caveats of open source and the confirmation that more data sometimes hurts machine learning models.
Seismic Compression and I/O
Equinor anounced the Python version of a compression algorithm for seismic data that supposedly speeds up i/o for seismic cubes by 8 times at a fixed bitrate. This should be particularly interesting for online learning of seismic data in machine learning and i/o intensive tasks such as seismic processing and inversion.
Find it on PyPI: https://pypi.org/project/seismic-zfp/
Why Open Source?
I am an avid listener of the Don’t Panic Geocast, a podcast about geoscience with great guests, interesting topics and the favourite segment fun-paper Friday. As I’m slow to catch up, I recently listened to the episode about Open Source. Why do we open source things? What is the business model? How does a highly profitable company like Apple rely on open source and contribute to it?
Find Episode 122 here: http://www.dontpanicgeocast.com/?p=477
Deep Double Descent
The double descent phenomenon has been known for a bit of time. It is the oddity of classical statistics and machine learning predictability not following the expected bias-variance trade-off. Open AI have confirmed this effect for very deep networks in accordance with the original research. Challenging the classic notion that “more data is always better”.