ACS61011 Deep Learning

Module Description (subject to change)

An important field within artificial intelligence is machine learning, which enables systems to learn from data rather than being explicitly programmed to solve a task. Conventional machine learning algorithms tend to rely on a human to carefully engineer and extract features to present to a machine learning algorithm, which can be time-consuming and difficult. A deep learning system, by contrast, takes raw data as input and learns to extract features automatically. This approach has led to significant improvements in processing images, video, speech and audio. Deep learning has also had an impact on the design of intelligent agents, giving rise to the area of deep reinforcement learning, which is where an agent learns in a reward-based framework. An example of deep reinforcement learning is where the Google DeepMind team designed an agent that learned to play Atari computer games to better-than-human-expert level.

This module will cover some of the key elements of deep learning including: the structure, training and evaluation of deep neural network classifiers and deep reinforcement learning systems, along with a view of advanced and state of the art methods. The module will be taught by a combination of lectures and labs, with an emphasis on the practical implementation of different algorithms.

Credits: 15 (Spring Semester)

Pre-Requisites: ACS6427 OR ACS61013

Module Leader

SAnderson

Dr Sean Anderson
Email: s.anderson@sheffield.ac.uk
Amy Johnson Building

If you have any questions about the module please talk to me during the lectures or the labs in the first instance. It is likely that other students will learn from any questions you ask as well, so don’t be afraid to ask.

Outside of lectures please contact me via email, or drop in to see me.

Learning Outcomes

Learning Outcomes

At the end of the module, the student should be able to:

1. Comprehend and explain deep learning, including motivation, problem formulation, algorithms, and future challenges.
2. Analyse and evaluate the design and implementation of deep learning methods.
3. Select and apply appropriate methods and computational tools to solve problems using deep learning.
4. Design and construct deep learning systems using appropriate methods.

This module satisfies the AHEP3 (Accreditation of Higher Education Programmes, Third Edition) Learning Outcomes that are listed in brackets after each learning outcome above. For further details on AHEP3 Learning Outcomes, see the downloads section of our accreditation webpage.

Syllabus

Syllabus 

Part 1 Introduction to Deep Learning (Based on Goodfellow et al., 2016, Deep Learning)
1. Introduction to deep learning
2. Machine learning basics
3. Deep feedforward networks
4. Convolutional networks
5. Optimisation
6. Regularisation

Part 2 Deep Reinforcement Learning (Based on Sutton and Barto, 2018, Reinforcement Learning: An Introduction)
7. Introduction to reinforcement learning
8. Reinforcement learning using tabular methods
9. Reinforcement learning using function approximation
10. Deep Q learning
11. Reinforcement learning using policy gradient methods
12. Deep deterministic policy gradient (DDPG)

Part 3 Advanced Methods (Based on Goodfellow et al., 2016, Deep Learning)
13. Recurrent neural networks
14. Deep recurrent networks
15. Generative adversarial networks
16. Future trends and challenges

Teaching Methods

Learning and Teaching Methods

NOTE: This summary of teaching methods is representative of a normal Semester. Owing to the ongoing disruption from Covid-19, the exact method of delivery will be different in 2020/21.

Lectures (16 hours, 2 hours per week) and problem classes (8 hours, 1 hour per week). Lectures will be used to introduce concepts in deep learning, and give guidance on practical work and assignments

Laboratory sessions (24 hours, 3 hours per week). Labs will be used for the design, practical implementation and evaluation of deep learning methods

Independent study (102 hours). Students are expected to spend time on independent study including directed reading, completing assignments and revision

Teaching Materials

Learning and Teaching Materials

All teaching materials will be available via Blackboard (MOLE).

Assessment

Assessment

4x Labs (5% each)

1x Open-ended Practical (20%)

Exam, 2hrs (60%)

No resit examination is available for this module.

Feedback

Feedback

Feedback will be given in the following forms:
•        Interactively during lab sessions.
•        Interactively during problem classes.
•        Individual feedback on lab work in the form of marked quizzes with worked solutions.
•        Individual feedback on the individual project in the form of solution evaluations and interactive feedback.

Student Evaluation

Student Evaluation

Students are encouraged to provide feedback during the module direct to the lecturer. Students will also have the opportunity to provide formal feedback via the Faculty of Engineering Student Evaluation Survey at the end of each semester.

You can view the latest Department response to the survey feedback here.

Recommended Reading

Recommended Reading

Goodfellow, I., Bengio, Y., Courville, A. (2016). Deep learning. Cambridge: MIT press

Sutton, R. S., & Barto, A. G. (2018). Reinforcement learning: An introduction. MIT press

Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A.A., Veness, J., Bellemare, M.G., Graves, A., Riedmiller, M., Fidjeland, A.K., Ostrovski, G. and Petersen, S., 2015. Human-level control through deep reinforcement learning. Nature, 518(7540), pp.529-533

Lillicrap, T.P., Hunt, J.J., Pritzel, A., Heess, N., Erez, T., Tassa, Y., Silver, D. and Wierstra, D., 2016. Continuous control with deep reinforcement learning. ICLR 2016. arXiv preprint arXiv:1509.02971