treora/toydata

Inspecting neural network internals using generated toy data

This repo contains the results of a course project I did for my machine learning studies. My idea was to try out generating toy data samples (20-bit vectors), and looking if the features that neural networks (RBM and MLP) would find would match the latent variables that were used to generate the data.

Reading the report

The project report is in the form of an IPython Notebook, allowing the reader to rerun the experiments while reading, and to modify them to try out new things.

To read and explore the report, clone the repo and run ipython notebook report/report.ipynb.

You will need:

• ipython,
• ipython-notebook
• the following python modules: (run pip install <module> to install modules)
  • numpy
  • matplotlib
  • sklearn
  • theano (only for MLP experiments)

Alternatively, you can take a read-only version of the report, as html or PDF.