An Introduction to Deep Learning
this sequence of lectures, I will give an elementary introduction to deep
learning, including deep neural network models, training algorithms and
applications. I will first briefly
discuss the concept of (multi-class) linearly separable sets and main ideas in
two classic linear models: support vector machine and logistic regression. I will then introduce a general function
class of deep neural networks and discuss their mathematical properties such as
linear independence, approximation properties and their relationship with the
classic finite element methods. As a
special deep neural network, I will discuss the convolutional neural network
(CNN) and its application to image classification. In particular, I will present a new class of
CNN framework, known as MgNet, which is motivated and derived from the classic
multigrid methods used for solving discretized partial differential
equations. If time allows, I will
discuss and analyze the commonly used training algorithm, stochastic gradient
decent method (SGD), and introduce some alternative new training algorithms,
known as extended regularized dual averaging (XRDA) method, that can be
effectively used for training sparse deep neural networks.
Part I: 14 June 2019 (Friday) from 9:30am to 12:00pm at Room G04, Liang Y C Hall, The Chinese University of Hong Kong.
Part II: 14 June 2019 (Friday) from 2:30pm - 17:00pm at Room G04, Liang Y C Hall, The Chinese University of Hong Kong.