Adding Titanic Tutorial #3

vivekmig · 2019-08-28T21:06:55Z

Adding tutorial for Captum based on the Titanic survival dataset. This tutorial reviews applying IG, Conductance and Neuron Conductance to a small model, demonstrating the different types of attributions possible using Captum. A few other small changes are included such as renaming Conductance to LayerConductance to make the distinction between LayerConductance and NeuronConductance clear as well as updating the default number of steps in conductance to 50 to match IG.

NarineK

Overall looks awesome! Thank you @vivekmig !

Couple comments:

What do you mean by the comment : # Linear 0 is simply identity transform did you mean linear1?
Also, I don't see relu-s in the forward function but there are some relu_out1 and relu_out2 variables.
Sigmoid is a little unusual did you mean Relu ?
It looks like the printouts for the training are very lengthy. Perhaps we could printout some of the epochs using an interval - after running for, let's say, each xx epochs, similar to CIFAR ?
nit: "how it reaches it's decision," -> "... its decision" ?
nit: "requiresgrad" -> requires_grad_ ?
Did you also try to compare with random forest's or the linear model's feature importance ?
Also, it would be interesting to compare the importances of different layers

vivekmig · 2019-08-29T18:52:57Z

Thanks for the feedback!

Whoops, this was left over from copying one of the test models, removed it now, thanks!
I was trying different models and switched ReLU to Sigmoid at some point and it was slightly better. Missed renaming the variable, so fixed that now, thanks!
Yeah, ReLU is more common, but Sigmoid did slightly better. I think it doesn't really affect the tutorial much either way, and I would need to update the later analysis if the network changes, so I'd prefer to leave it to Sigmoid.
Good point, I've updated to only print every 20 epochs now.
Fixed, thanks!
Seems like this was markdown interpreting underscore for italics, fixed this now, thanks!
So I did try with logistic regression initially and originally had the tutorial starting with the logistic regression model and then developing to the NN models. Overall the importances are pretty similar, but the issue with that was that the age coefficient looked artificially very small, because the range was much larger. I could've normalized the variables initially, but it was also nice to have the 0/1 significance for binary variables. Also, the normalization isn't really necessary for IG, so I didn't want to make it more confusing than necessary. I finally decided that the tutorial was reasonably long already and the logistic regression didn't seem to add much value to conveying the usage / importance of Captum (and may actually make it a little more confusing), so I ended up removing it. If you think it would be valuable to add back, lmk, I can add it!
How do we define the importance of a layer / compare across layers? I could do an analysis of the second layer as well, but I figured the first layer might be sufficient to show the usage.

NarineK · 2019-08-29T19:46:04Z

Thank you, Vivek!
7. For the comparison, we can think later creating separate notebook. Titanic DS is usually analyzed with traditional ML Algorithms. I was thinking that showing what those models learn differently and what features does one model focuses on vs the other, would be useful.
8. I was thinking of comparing the importances of those layers (e.g. comparing the sum of importances of the layers and the distribution of importances). E.g. if second layer looks exactly the same as the first one - is it more important ? This will make sense with more layers and would probably have more scientific focus rather than "how to use captum".

It looks good for showcasing how to use captum.

orionr

This is awesome! Some nits, but approving regardless.

orionr · 2019-08-29T20:04:50Z