MSc Machine Learning

A key property of real-world data is that it is often presented in large dimensions and yet many of its relevant properties can be encoded in much lower dimensions. The ability to find such low dimensional representation of data can hugely aid the feasibility of the downstream Machine Learning (ML) tasks that one might want to accomplish using it. Further, real-world data is often unlabelled. In this module we will learn to discover such efficient representations of unlabelled data and learn how it can be leveraged to extract insights. 

There are two primary aims of this module (a) to introduce the students to thinking systematically about unsupervised machine learning with high-dimensional data and (b) to get the students become familiar with implementing on modern software some of the classical algorithms of this type. 

1. Describe the fundamental concepts of unsupervised learning and in particular the methods of detecting effective low dimensional descriptions of data, if such exist.  
    
2. Analyze the differences between the methods and utility of various ML methods that detect effective low dimensional structure in data.
    
3. Explain the models and training algorithms used to implement a (Variational)Auto-Encoder (V)AE. 
    
4. Apply different dimension reduction methods in a judicious way depending on the data and the need. 
    
5. Implement different unsupervised ML methods in standard ML-specific software. 
   
    

• Principal Component Analysis (and its variants)
• Spectral Clustering
• Dimension Reduction  
• Boltzmann Machines
• (Variational) Auto-Encoders  

• Asynchronous material in the form of video lectures, formative exercises, lecture slides and code examples delivered via the virtual learning environment, lectures and supported tutorials.  
   
• Weekly lectures consolidating asynchronous materials and providing opportunities for discussion and questions.  
   
• Weekly supervised lab sessions provide support for coursework and formative exercises.   

Other: Coursework (50%) 
 

• Individual feedback would be available on request when coursework scores are released.  
   
• Cohort level feedback will be available for the final exam, after marking.  

“Introduction to Statistical Learning”, 
G. James, D. Witten, T. Hastie, R. Tibshirani, Springer, 2023.  
 

“Probabilistic Machine Learning: Advanced Topics”,  
K.P. Murphy, MIT Press, 2023. 

Students are expected to spend 20 hours on Coursework.