On the Identifiability of Switching Dynamical Systems

We are delighted to present our CHAI Seminar Series, organised by our Early Career Researchers (ECRs). This seminar series will provide a platform for knowledge exchange and discussion on cutting-edge research in causal AI and related fields.

Speaker : Dr Yingzhen Li

Dr. Yingzhen Li is currently a Senior Lecturer in Machine Learning at the Department of Computing at Imperial College London. Her passion lies in creating reliable machine learning systems capable of generalizing to unseen environments, combining the strengths of Bayesian statistics and deep learning to achieve this goal.

Dr. Li's research interests are twofold, rooted in the application of probabilistic machine learning methods:

1. Trustworthy Machine Learning Models: She focuses on key areas such as uncertainty quantification, robustness, explainable AI, decision-making, and adaptive methods, including continual learning and model editing.

2. Generative Modelling: Her work extends to sequential generative models for a variety of data types, including video, spatiotemporal data, general time-series, images, and tabular data. Additionally, she is interested in causal representation learning using generative models.

A significant portion of Dr. Li's work has been dedicated to approximate inference, with applications in Bayesian deep learning and deep generative models. These contributions have not only advanced industrial systems but have also been integrated into popular deep learning frameworks, such as TensorFlow Probability and Pyro. For an in-depth understanding of her expertise, refer to her tutorial on approximate inference presented at NeurIPS 2020.

Before joining Imperial College London, Dr. Li was a Senior Researcher at Microsoft Research Cambridge and previously interned at Disney Research. She earned her PhD in Engineering from the University of Cambridge, UK.

On the Identifiability of Switching Dynamical Systems

One of my research dreams is to build a high-resolution video generation model that enables granularity controls in e.g., the scene appearance and the interactions between objects. I tried, and then realised the need of me inventing deep learning tricks for this goal is due to the issue of non-identifiability in my sequential deep generative models. In this talk I will discuss our research towards developing identifiable deep generative models in sequence modelling, and share some recent and on-going works regarding switching dynamic models. In particular, we first show conditions of identifiability for Markov Switching Models (or auto-regressive HMMs) with non-linear transitions, with a new proof technique different from the algebraic approach of the seminal HMM identifiability work by Allman et al. 2009. Then we lift the Markov Switching Model to latent space and leverage existing results to show identifiability. If time permits, I will also show recent developments that build in more flexible structures in the latent switching dynamical prior.

Please note that this seminar will be recorded. If you would like to opt-out of the recording or have any questions about this, please email chai@ed.ac.uk