UNIT INFO

COMSM0045 - Applied Deep Learning

ADL Banner

Unit Information

Welcome to COMSM0045. The unit introduces the students to deep architectures for learning linear and non-linear transformations of big data towards tasks such as classification and regression. The unit paves the path from understanding the fundamentals of convolutional and recurrent neural networks through to training and optimisation as well as evaluation of learnt outcomes. The unit's approach is hands-on, focusing on the 'how-to' while covering the basic theoretical foundations. For further general information, see the syllabus for the unit.


Staff

Dima Damen (DD)Unit Director
Tilo Burghardt (TB)

Teaching Assistants

Hazel Doughty (HD), Will Price (WP), Evangelos Kazakos (EK), Jonathan Munro (JoM), Jian Ma (JiM), Xinyu Yang (XY), Dan Whettam (DW), Adriano Fragomeni (AF)


Unit Materials

Wks Pre-Recorded Lectures Friday Synch Session
10am-1pm
Labs
0 Wk0
INTRODUCTION TO COMSM0045 (Video)
GETTING STARTED:

Register Individually on BlueCrystal4
(details see below)
1 Wk1 - LECTURE 1
BASICS OF ARTIFICIAL NEURAL NETWORKS
(Live on Teams at 10am 09/10/20, Slides)
(Introduction, Neural Networks, Perceptron, Cost Functions, Gradient Descent, Delta Rule, Deep Networks)

Wk1 - LECTURE 2
TOWARDS TRAINING DEEP FORWARD NETWORKS
(Live on Teams at 11am 09/10/20, Slides)
(Network Representation, Computational Graphs, Reverse Auto-Differentiation)
(no scheduled lab for week 1)
RECAP WORKSHEETS:
-Convolutions (Homework)
-Lab0 - Python (Homework)
2 Wk2- LECTURE 3
BACKPROPAGATION ALGORITHM
(Video, Slides)
(The Backpropagation Algorithm in Full Detail, Activation Functions)

Wk2 - LECTURE 4
OPTIMISATION TECHNIQUES
(Video, Slides)
(Stochastic Gradient Descent, Nesterov Momentum, RMSProp, Newton's Method, AdaGrad, Adam, Saddle Points)
16/10/20,10am,Online - PRACTICAL 1 (Slides), (Video)
Your first fully connected layer
gradient descent
stochastic gradient descent
16/10/20, Online - 3hrs
-BC4 Setup
Lab 1 - Training your first Deep Neural Network
3 Wk3 - LECTURE 5
COST FUNCTIONS, REGULARISATION AND DEPTH
(Video, Slides)
(SoftMax, Cross Entropy, Hingeloss, L1 and L2 Regularisation, DropOut, DropConnect, Depth Considerations)
23/10/20,10am,Online - PRACTICAL 2 (Slides), (Video)
Your first convolutional connected layer
23/10/20, Online - 3hr
Lab 2 - Your First Convolutional Connected Network
Wk3 - LECTURE 6
CONVOLUTIONAL NEURAL NETWORKS
(Video Part 1, Video Part 2, Slides Part 1, Slides Part 2)
(sharing parameters, conv layers, pooling, CNN architectures)
4 - 30/10/20,10am,Online - PRACTICAL 3 (Slides), (Video)
Error rate monitoring (training/validation/testing)
Batch-based training
Learning rate
Batch normalisation
Parameter intialisation
30/10/20, Online - 3hr

Lab 3 - Hyperparameters
5 - 6/11/20, 10am, Online
Continuation Lab
6/11/20, Online - 3hr

Lab
Continuation
6 - 13/11/20, 11am, Online - Practical 4 Intro (Slides) (Video)

PRACTICAL 4
Data Augmentation
Debugging strategies
Dropout
13/11/20, Online - 3hr
Lab 4
Data Augmentation
7 Wk7 - LECTURE 7
RECURRENT NEURAL NETWORKS
(temporal dependencies, RNN, bi-directional RNNs, encoder-decoder, LSTM, gated RNN)
(Slides)
20/11/20, 10am, Online
Continuation Lab + First CW Lab
20/11/20, Online - 3hr

Lab
Continuation + Project Start
8 - 27/11/20, 10am [1 hour], Online - CW Q/A -
9 - 4/12/20, 10am [1 hour], Online - CW Q/A -
10 CW Deadline Fri 11 Dec 13:00 (Blackboard Submission)
11 - 18/12/20, 10am [1 hour], Online - Exam Q/A-

Assessment Details

All students in the unit are requested to study the paper: Pan et al (2016). Shallow and Deep Convolutional Networks for Saliency Prediction. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Available on ArXiv

Coursework specs are now available at: COMSM0045-COURSEWORK-SPECS-2020


Assessment Details - Exam

Open book 2 hours exam in January (online). See note above on additional reading for exam.


Github

All technical resources will be posted on the COMSM0045 ADL Github organisation. If you find any issues, please kindly raise an issue in the respective repository.


Textbook

Recommended Reading:
Goodfellow et al (2016). Deep Learning. MIT Press


Blue Crystal 4 Registration [only applicable for Bristol undergraduate students with corresponding email]

All students must apply online to register an account on BC4 for this unit. This also applies to students who already have accounts on BC4 for other units (e.g. HPC), in this case you must register again using the instructions below.

  1. Click on: https://www.acrc.bris.ac.uk/login-area/apply.cgi
  2. Enter your personal details
  3. Choose: "Join an existing project"
  4. Enter project code: COSC020582
  5. Keep Preferred log-in shell as bash
  6. Do not provide any additional information

Note that it takes up to 48 hours to enable your account on BC4.