Ensemble

After training of all the networks we needed to make a decision: Which networks are worthy to be integrated in the final classifier regarding accuracy but also regarding the amount of computation time.

We choose only the following two networks:

• the CNN for README
• and the CNN for the Commit times.

These give us their predictions and we feed them into our master net, the Ensemble Network.

This is an easy two-layer feed forward topology for learning non-linear weightings of the CNN predictions. For example it should learn to rely only on the Commits if the README is empty.