Deep Learning Seminar

"Universal sparsity of deep ReLU networks"

Elbrächter, Dennis

In recent years Deep Learning has been successfully applied to a
variety of very different problems too numerous to fit on a single
page. While this might very well constitute its most attractive feature
in practice, understanding why it is so universally useful remains
a compelling challenge. We try to approach the issue from
a sparsity point of view which leads to a remarkable approximation
theoretic universality property of deep neural networks. We introduce
(or assimilate) a number of key concepts, which allows us to compare
neural networks to classical representation systems (meaning e.g.
wavelets, shearlets, and Gabor systems, or more generally any system
generated from some mother function through translation, dilation
and modulation). This enables us to establish that any function class
is (asymptotically) at least as sparse w.r.t. (ReLU) neural networks,
as it is in any ’reasonable’ classical representation system.

« back