Neural Network examples and experiments performed during the summer of 2016. The first 2 sections are illustrative of how neural networks work in general, while on the last 2 I create a RNN trained on Charles Bukowski's poems to write text that look like his. All is done on Numpy and Theano, following the excellent WildML's tutorials.
-
Numpy Network: Base numpy-based implementation of a neural network for 2D classification.
-
Theano Network: Same problem as above, but using a Theano implementation instead.
-
Numpy Bukowski: Recurrent neural network (RNN) using Numpy. Here I start using a collection of Bukowski's poems as training data.
-
Theano Bukowski: Reimplementation of the Bukowski neural network, but now using Theano, as the base Numpy model takes significantly longer to run. Sample output poems are available here.
-
LSTM Bukowski: Final attempt to generate more coherent poems, using a LSTM neural network. Currently under construction.
- Sentiment Bukowski: Sentiment analysis on the same set of poems used to build the neural networks. I look into the most common subjects occuring on the text, as well as associations the author makes when he is talking about men and women.