Courses Offered

COMP3340 Applied Deep Learning

2021-22

Instructor(s):

Luo Ping

(Class A)

No. of credit(s):

Recommended Learning Hours:

Lecture:	26.0
Tutorial:	13.0

Pre-requisite(s):

COMP2119 or ELEC2543 or FITE2000; and MATH1853

Co-requisite(s):

Mutually exclusive with:

ELEC4544

Remarks:

Course Learning Outcomes

1.	[1] understand the motivations and principles for building deep learning systems based on empirical data, and how deep learning relates to the broader field of artificial intelligence.
2.	[2] formulate problems associated with domain specific data (e.g., image recognition, image generation, reinforcement learning, and language translation) in terms of deep learning models.
3.	[3] implement solutions to computer vision, natural language processing, and robotic problems using deep learning toolboxes such as PyTorch or Tensorflow, apply numerical optimization algorithms.

T - Teach, P - Practice
For BEng(CompSc) Programme Learning Outcomes, please refer to here.

Syllabus

Calendar Entry:

Detailed Description:

Introduction	Mapped to CLOs
Course Introduction Overview of computer vision, natural language processing, reinforcement learning in video game, and historical context. Introduction to Python/NumPy.	1
Loss Functions and Optimization Linear classification, higher-level representations, image features, optimization, stochastic gradient descent (SGD).	1, 2
Introduction to Neural Networks Backpropagation, Multi-layer Perceptrons.	1, 2
Case #1: Able to build neural networks with backpropagation and SGD.	2, 3
Convolutional Neural Networks	Mapped to CLOs
Convolutional Neural Networks History, Convolution and pooling, ConvNets beyong vision, Introduction to PyTorch	1, 2
CNN Architectures AlexNet, VGG, GoogLeNet, ResNet, etc	1, 2
Case #2: Able to finetune and evaluate deep neural networks (e.g. AlexNet) for image recognition and others.	2, 3
Generative Adversarial Networks	Mapped to CLOs
GAN Architectures Vanilla GAN, WGAN, DCGAN, CGAN, BEGAN, SRGAN, etc.	1, 2
Case #3: Able to finetune and evaluate SRGAN on CelebA for face image generation, super resolution, and others	2, 3
Reinforcement Learning	Mapped to CLOs
RL basics: policy gradients, hard attention	1, 2
Q-Learning	1, 2
Actor-Critic	1, 2
Case #4: Able to train CNN_DQN to play Atari.	2, 3

Assessment:
Continuous Assessment: 50%
Written Examination: 50%

Teaching Plan

Please refer to the corresponding Moodle course.

Moodle Course(s)