Lesson 12 AMA

@jeremy Will we be studying the following in the last class of part 2 ?

  1. Siamese and Triplet networks.
  2. Applying Deep learning for Database table data (Structured).
  3. How to Productionalize DL models.

Hey Jeremy,

Just wanted to know about productionizing DL models. Would we be having a little session on that? If not could you point out to some specific things to read up on? Any particular papers or tutorials that you suggest going through?


Has anyone come across the “Universal adversarial perturbations” paper (https://arxiv.org/abs/1610.08401)?

It looks to me that it could have interesting applications in privacy. Any thoughts?

I am also interested to know this. I am currently working as Software Engineer (past 2 years after undergrad) in a product company in India. Seems like we have a machine learning labs division. This course has encouraged me to ask for team change etc, But don’t know to what extent I will be able to apply deep learning there … And the peer group is very less. I mean there are not many good research work/open source dev etc in India pertaining to deep learning that I know off.

The opportunities are very less compared to that in valley.

TLDR; What things should I do, as a engineer outside united states/europe can I join deep learning companies doing cool stuff after this course?

I believe some companies in the bay area have remote teams - might that be an option? If so, perhaps you could transfer to the US after a year or so. I agree that it’s really important to have a peer group.

I want to know this too. Haven’t start part 2(studying part 1), do lessons of part 2 show us how to use keras models with minimal dependencies?

Keras api is well designed, examples are easy to read and understand, but to create a stand alone app by keras is another problem. Is it possible to generate cross platform c++/c api with minimal dependencies?

Would lessons of part 2(or part 3?) introduce us how to do that?Thanks

ps : dlib and tiny-cnn are easy to compile, but they lack a lot of features when comparing with keras, right now I do not have the ability to implement all of those missing layers for those libraries yet(I tried to develop a pixel wise mean squared error loss function for super resolution, but fail).

Hi, I would like to know when you found Enlitic the AI imaging company for lung cancer, what was your business model, why did you think it will work? how did you execute your plan?

Hi @jeremy I listened to a podcast on Cardiologist-Level Arrhythmia Detection with Convolutional Neural Networks https://dataskeptic.com/blog/episodes/2017/cardiologist-level-arrhythmia-detection-with-cnns
Very clear paper: https://arxiv.org/pdf/1707.01836.pdf
A CNN is used to solve a sequence to sequence problem. The author stated in the podcast (at 22:50) that a CNN was preferred over an RNN, because CNNs are significantly faster to train. Can you comment on this design choice?