autoenc-neurons (work in progress)

An experiment with convolutional autoencoders as applied to text, and exploring what the 'neurons' in the convolutional layers respond to.

Running

Note that if you do not have GPUs, you will need to edit the script to install tensorflow instead of tensorflow-gpu, and change gpu_count to 0.

python3 autoenc-neurons.py

This will start model training and then produce output, at last, like this:

The output contains several lines from the text, repeated several times. In each stanza, the letters are shaded according to the activation that it causes a particular neuron (darker means more active). This is repeated for several neurons.

References

The source text is The Complete Project Gutenberg Works of Jane Austen:
http://www.gutenberg.org/files/31100/31100.txt

For more on convolutional autoencoders in Keras, see: https://blog.keras.io/building-autoencoders-in-keras.html