杏吧论坛

Prof Milan Vojnovic COL 5.05

This course is available on the MSc in Applied Social Data Science, MSc in Data Science, MSc in Quantitative Methods for Risk Management, MSc in Statistics, MSc in Statistics (Financial Statistics), MSc in Statistics (Financial Statistics) (杏吧论坛 and Fudan), MSc in Statistics (Financial Statistics) (Research), MSc in Statistics (Research), MSc in Statistics (Social Statistics) and MSc in Statistics (Social Statistics) (Research). This course is available with permission as an outside option to students on other programmes where regulations permit.

The course will introduce the basic principles and algorithms used in artificial intelligence systems, with a focus on statistical and machine learning foundations, including the design and training of deep neural networks and reinforcement learning algorithms. These algorithms will be applied to classification tasks such as image recognition, speech recognition, natural language processing, as well as to agents learning to play various games. Use of Python and open source software libraries for machine intelligence such as Tensorflow and OpenAI Gym constitutes an integral part of the course, allowing students to gain hands-on experience in data analysis and use of modern computational tools.

20 hours of lectures and 15 hours of computer workshops in the LT.

Week 6 will be a reading week.

The course will start with an introduction to basic concepts of artificial intelligence systems that include supervised, unsupervised, and reinforcement learning. It will then cover basic concepts of deep neural network models covering the topics of feed forward networks, perceptron, learning XOR function, architecture of deep neural networks and the concept of hidden units. The course will also cover the topic of optimisation for deep learning including concepts such as stochastic gradient descent methods, backpropagation, parameter norm penalties, dataset augmentation, early stopping and dropout. Special architectures of deep neural networks will be studied in more depth including convolutional neural networks, explaining the intuition underlying their design, the concept of pooling and efficient convolution algorithms. Other architectures will be studied as well including those used for sequence modelling, such as recurrent neural networks, the long-short term memory and other gated recurrent neural networks. The course will cover the main principles of reinforcement learning algorithms, explaining the main concepts such as rewards and punishments, exploration versus exploitation, Markov decision processes frameworks, and various solutions methods.

Examples and use cases will be drawn from various application domains including classical applications in the domains of computer vision, speech recognition, natural language processing and strategic game playing. Students will have an option to study application of deep learning using a dataset in the application domain of their interest including those in the area of social and economic systems.

The course will be based on using modern software frameworks for learning deep neural networks including open source Tensorflow and OpenAI Gym software libraries. Students will develop their code using Jupyter / Google Colab notebooks. The course will use GitHub platform for maintaining a repository of lecture material, homework and project assignments and anything related to the running of the course. The course handout is available here https://lse-st449.github.io/.

Students will be expected to produce 10 problem sets in the LT.

• I. Goodfellow, Y. Bengio and A. Courville, Deep Learning, MIT Press, 2017, http://www.deeplearningbook.org 
• R. Sutton and A. C. Barto, Reinforcement Learning: An Introduction, Second Edition, MIT Press, 2018
• TensorFlow, An Open Source Software Library for Machine Intelligence, http://www.tensorflow.org
• OpenAI Gym, https://gym.openai.com/

Project (80%) in the LT.
Continuous assessment (10%) in the LT Week 4.
Continuous assessment (10%) in the LT Week 7.

Two problems sets submitted by students will be assessed (20% in total). In addition, there will be a graded take-home research project (80%) which will completed on an individual basis in which they will demonstrate the ability to apply and train an appropriate model to a specific problem and dataset using principles they have learnt in the course. This may be done by publishing the code to a GitHub repository and GitHub pages website.