16 comments

[ 3.2 ms ] story [ 26.0 ms ] thread
Me and the other authors of the post will be around if anyone has any questions.
I'm doing some research into handwriting recognition. In particular transcribing older documents. Do you see a way of applying your work here in that direction?
The generative model of handwriting that we're working with here probably isn't very applicable to handwriting recognition.

In principle, you could train a model like this to jointly model the text and the produced handwriting and then search for the most likely text to correspond to handwriting, but it would be a lot less efficient and likely not work as well. Instead, the natural way to apply neural networks to use a convolutional neural network. You could either predict the presence of characters at different positions and stitch them together with another program, or do an end-to-end image to sequence approach, probably using attention.

If you want to visualize that kind of model, the techniques you want to use are pretty different than what we have in this article. But there are some pretty useful techniques! In particular, you could do attention visualization to understand where your model is looking as it predicts particular characters and optimization-based feature visualization to understand what different features in your model represent.

You can still have CNN hallucinate most likely set of next inputs.

And the kind of activity display you did later also works.

why does it such so much and constantly assume i'll write another "c" character?

i don't see anything worth anything.

How much computing power/time did it take to train the model? Had any hyperparameters optimization been done?

Sorry if I missed the reference that has this information, but I think that these numbers are important in any deep learning experiment because they allow the readers to evaluate applicability of described methods to their problems.

Quite interesting to see the inspection of the model.
We're slowly coming to an era where everything can be easily forged.
Except digital signatures. Those are still hard and much easier to steal or phish than crack.
This is incredible important research. There are fields in which adoption of ML techniques is dependent on it being explainable and inspectable to stakeholders, e.g. in certain health care, policy or finance applications.

There is work being done on explaining non-linear models, such as LIME [1], but much more is needed. At a recent d3 meetup the topic was discussed but we primarily noted its importance. Notes at [2].

[1] https://github.com/marcotcr/lime [2] https://docs.google.com/document/d/1o-DO65PEZShLuadyRd35PsTJ...

In "Examining the Internals of the Model" it could be interesting to show an overview activation map that sums over the absolute activation of all the examples—to highlight which cells activate on all examples—and one that is colored by variance of activation across examples—to show which cells influence the character/differences in writing.