rando

about rando

rando is a demo of the power of recurrent neural networks in natural language processing, especialluy multi-layer (deep) networks.

For this mini app, a so-called long short-term memory neural network was trained on the titles of more than 9000 papers in SEG's GEOPHYSICS journal, and another was trained on their authors. The LSTM code was adapted from Andrej Karpathy's char-rnn project. This app samples from the network, seeded by a random word stem. The result is a (mostly) nonsensical pseudo-language that looks a bit like geophysics.

credits

This project was built for fun by Matt Hall of Agile Scientific (email Matt). You can read more about it in Hall, M (2017), Three data analytics party tricks, The Leading Edge, March 2017. It's part of a series of experiments hosted at geosci.ai.

This web app is built with Lua/Torch and Python's Flask framework, and is hosted on an Amazon EC2 server.

Thank you to SEG and especially Isaac Farley for permission to use this dataset.