HTC Seminar Series #19- The Epistemology of Deep Learning

 Today's HTC Seminar Series presentation is a talk by Yann LeCunn (Facebook, New York University) at the Institute for the Advanced Study (IAS) in 2019 on The Epistemology of Deep Learning.


This talk was from a conference titled 'Deep Learning : Alchemy or Science'?


'Epistemology' is a big word.  If you look it up you get something along the line of 'the theory of knowledge, especially with regard to its methods, validity, and scope'.  So 'Epistemology' is the investigation of what distinguishes justified belief from opinion.

And that's probably worth taking a look at closely in this area of research and implementation we call 'deep learning'.

So i looked up 'Alchemy', and got a lot that doesn't really apply to this blog, but the one i'm going to pass on is the following  'a seemingly magical process of transformation, creation, or combination.'  

Now if that sounds like a good description of a magical GAN deep learning architecture to you (and it certainly does to me), then maybe there is a purpose to this particular conference.  And maybe we are performing some kind of magical deep learning alchemy.  Now is that because we're smart, or because we're stupid?  The fact that it seems so magical?

Certainly we need to look no further than how fastai keeps discovering things that appear to defy conventional wisdom in the field, and then next year become one of the things you put in your bag of best practices.  U-Nets, anyone?  How about actually plotting the performance of the learning rate, and picking the best value to start with for training?  We could continue on with this list, but i think you get the idea.


Comments

Popular posts from this blog

CycleGAN: a GAN architecture for learning unpaired image to image transformations

Pix2Pix: a GAN architecture for image to image transformation

Smart Fabrics