HTC Seminar Series #36: Mathematical Mysteries of Deep Neural Networks

Today's HTC Seminar Series presentation is a lecture by Stephane Mallat at the ICTP Colloquium recorded in November 2020. He cuts through the BS and gets to the real meat of what is going on under the hood in deep learning neural nets.

ABSTRACT: Deep neural networks obtain impressive results for image, sound and language recognition or to address complex problems in physics. They are partly responsible of the renewal of artificial intelligence. Yet, we do not understand why they can work so well and why they fail sometimes, which raises many problems of robustness and explainability. 

Recognizing or classifying data amounts to approximate phenomena which depend on a very large number of variables. The combinatorial explosion of possibilities makes it potentially impossible to solve. One can learn from data only if the problem is highly structured. Deep neural networks appear to take advantage of these unknown structures. Understanding this "architecture of complexity" involves many branches of mathematics and is related to open questions in physics. I will discuss some approaches and show applications.


Observations:

1:  Today's lecture really does give you the keys to the deep learning kingdom.  Too bad it has only 1600 views.  The rest of the world's loss, and your gain.

2:  Think about what he is saying and how it relates to group theory and perceptual manifolds.

3:  Something that might be missing from this approach is a data augmentation fold in. Data augmentation is really a key component of deep learning on conventional CNN architectures.  How does the concept fit in here?

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