I’m sharing a PyTorch implementation of a neural turing machine.
I’m well aware that there are a few other implementations out there, but I think mine is quite straight forward yet extensible/generic, plus it works great. This is the result of intensive 3 weeks. Also, I documented anything worth documenting so it’ll be easy to pick up and use in external projects.
Here’s an animated GIF that shows how the network learns to predict the targets for the repeat-copy task. Specifically, the network was evaluated in each checkpoint that was saved during training (with the same input sequence).