Not Just a Black Box: Learning Important Features Through Propagating Activation Differences

[agitter]

https://arxiv.org/abs/1605.01713

At a glance: A strategy for interpreting neural networks, which could be an important topic in the review. The paper is a technical description of a general method, but one subsection shows an application to genomics and there is more work coming from the Kundaje lab in this direction.

[akundaje]

We'll have an expanded version of this on Biorxiv in the next few weeks.

[cgreene]

@akundaje : Will you post here when it's up? Might put this one on hold & read the new one at the same time.

[akundaje]

@cgreene Yeah definitely. The current version is extremely succint (had to be that way due to conference format limitations). So the new expanded version will be much easier to read and understand.

[michaelmhoffman]

@akundaje: I just came across your DragoNN web site: http://kundajelab.github.io/dragonn/

Is this the same thing? Or is there another paper for that?

[cgreene]

I saw @akundaje give a talk at the Simons Network Biology workshop. It may be
some of that work:
https://simons.berkeley.edu/talks-anshul-kundaje-04-13-16

Side note/shameless plug since I also gave a talk - there's actually a lot
of very nice talks. For all of the ones where the speaker gave permission,
the talks are recorded and available on youtube:
https://simons.berkeley.edu/workshops/schedule/1806

Edit: actually tag people's github names for ease of use.

On Tue, Aug 9, 2016 at 8:16 AM Michael Hoffman notifications@github.com
wrote:

@akundaje https://github.com/akundaje: I just came across your DragoNN
web site: http://kundajelab.github.io/dragonn/

Is this the same thing? Or is there another paper for that?

—
You are receiving this because you were mentioned.

Reply to this email directly, view it on GitHub
#50 (comment),
or mute the thread
https://github.com/notifications/unsubscribe-auth/AAhHs7iwlg9eabKF_I0D-8KYYf93KASlks5qeG-EgaJpZM4JeC43
.

[akundaje]

@michaelmhoffman We have a different paper in submission for DRAGONN. It'll be up on biorxiv by the weekend. DeepLIFT is provided in DRAGONN for identifying predictive features in the input sequences.

[akundaje]

@cgreene Yeah the Simon's meeting was really great.

[jbloom22]

Here is Anshul's Nov 2 presentation on this and related work from the Models, Inference & Algorithms seminar at the Broad Institute.

[akundaje]

@cgreene @agitter Here is our significantly expanded and updated preprint on DeepLIFT https://arxiv.org/abs/1704.02685 . Submitted to an ML conference. @AvantiShri made some great videos explaining DeepLIFT here http://goo.gl/qKb7pL . Updated code: http://goo.gl/RM8jvH.

[akundaje]

Also @cgreene @agitter we will finish a draft of the interpretation section by next week.

[agitter]

Thanks @akundaje. @jacklanchantin can you please see if there are changes you'd like to incorporate into your DeepLIFT description?

Regarding interpretation, @blengerich wrote something in #312. I propose moving that to the Discussion so you can use his text as a starting point. I'll make a pull request.

[agitter]

The updated version is covered in the Discussion

[agitter]

@traversc noted that this was published in the ICML 2017 proceedings. There is no DOI as far as I can tell:
http://proceedings.mlr.press/v70/shrikumar17a.html

[AvantiShri]

That appears to be correct. I suspect PMLR didn't think it needed to purchase DOIs since they have their own identifier system (PMLR 70:3145-3153 in this case).