Hi Ian,
I’m researching at the cross section of AI and Law and have a few questions for which I would appreciate your thoughts. These questions are not exactly about Deep Learning but at the fundamentals of Deep Neural Networks. Questions:
Do we have the ability/ need to interpret the the synoptic weights in Neural Networks? To a layman as well as to expert machine learning researchers, neural network weight matrices don’t convey any meaningful information. I believe there is some ongoing research in algebraic topology to understand what mathematical spaces these un-interpretable weight matrices occupy. What is your intuitive understanding on the interpretability of Neural weights (connection between the neurons)? The reason for asking this question is to understand the mutation of input till the predicted output (Since, neural net weights don’t convey any actionable information on the influence of input predictors on the outcome)
Is it time again (a very strong debate around this topic was active in 1990s) that we ask for copyright on Neural weights (Does Google copyright the neural weights of its ML platforms)? One of the secret saucages behind how effective a Deep neural network is the weights arrived after thousands of hours worth of training. Seeing the time and investment that went behind arriving on those weighted numbers, it is indeed very valuable. Your thoughts?
I am trying to use GANs for generating medical data. I am planning to give correlation between features as an extra information to the GAN (think of it as a knowledge based model) but not sure how to incorporate explicit correlation based loss along with the original loss function. Any tips and insights?
Thanks.
If tech support is on here , the webinar is not working. Specifically, I try to register, hit submit button and nothing. I tested another webinar and it works just fine, so it is not my setup.
I believe that @Negar_Rostamzadeh maybe will be able to help since she was the moderator of this ACM Webinar. Negar, this ACM Webinar is no longer viewable, me and @Willy_Rempel have tried to register to see the archived webinar but it does not work (but it works for other past webinars). Do you know who can fix it?