Introduction to Representation learning: Approaches, Challenges and Applications

Speaker :
Shuyu Lin
University of Oxford

Abstract:
Representation learning is one of the central topics in today's machine learning and computer vision research. However, the concept of a representation is often vaguely defined.
In this talk, I will start with clarifying these questions: what is the representation that we want our AI algorithms to learn? What makes a good representation? I will then introduce an unsupervised learning approach (variational autoencoders (VAEs)) that many researchers use to derive a representation for images and audios. We will discuss the strengths and limitations of this approach and some research that I have carried out in order to improve the representations derived from VAEs. We will end the talk by looking at a few applications (anomaly detection and data manipulation) which are enabled by the representation learning.
I hope this talk will explain why representation learning is important for many AI applications and motivate you to consider doing research in the related topics!


Speaker Bio:
I am a 3rd year PhD student at Cyber Physical Systems Group in Computer Science Department, University of Oxford. During my PhD, I aim to develop algorithms which can truly 'think’ and 'understand’. This requires algorithms to be more capable than solving specific tasks. They should be able to form meaningful internal representation of the complex raw data that they observe and ready to use the derived representation for future tasks. Concretely, my research involves studying existing representation learning algorithms, proposing new representation learning algorithms and deploying these algorithms on different application scenarios.
Outside research, I care about science communication, the environment and am very keen on promoting diversity & equality.

To attend more talks and interact with the Researchers follow us on :
Twitter:   / talks_cv  
LinkedIn :   / comp.  .
Website: https://computervisiontalks.github.io