Table of Contents
Going beyond the encoder-decoder architectures of RNNs, I’m experimenting with the attention model.
The original paper by Bahdanau et al is quite interesting, and attention seems to be all the rage right now.
This experiment takes as input an English sentence
five minutes and seven seconds past noon and predicts the time
The details can be seen in the notebooks, one for the model and another for the attention map.
I think the illustration suffers a bit due to the short inputs, so I’ll be working on an experiment that uses longer inputs (perhaps revisiting my fake news generator).
Also, I was again reminded that most of the work goes into preparing the dataset. I spent a considerable amount of time on the human/machine time generator (and associated test cases).
This graph shows where the neural net “looked” at
five minutes and seven seconds past noon when piecing together
12:05:07. The word
seven has a clear attraction to
07. As for the hours and minutes, fellow ML-er Neil thinks the saliency is spread out more for the minutes and hours, due to the net accepting “to” vs “from”, and so forth.
There you have it, ladies and gents. Amateur hour is over, and I’m excited to work on my next small showcase (whilst chiseling away at the new model for my side project).