Prof Ramamoorthy talking at Φ-ML meets Engineering
3rd June at 1pm
As part of the Φ-ML meets Engineering seminar series, Prof Ramamoorthy will give an online talk on 3rd June at 1pm. This talk, organised through the Alan Turing Institute, will be on "Online and hybrid system identification directly from raw sensory signals".
Φ-ML meets Engineering is a bi-monthly (online) seminar series discussing applications of Physics-enhanced Machine Learning methods in Engineering practice.
Details on how to subscribe to the seminar series to join this talk can be found here.
Talk Title: Online and hybrid system identification directly from raw sensory signals
Abstract: Adaptive systems in engineering depend crucially on identification of the underlying system from sparse observations. Similar problems arise in the experimental sciences in the form of inverse problems. With rapid advances in fields such as robotics and other autonomous systems, it becomes increasingly more important to find efficient ways to solve these problems, going directly from rich sensory feeds to structured models. We present recent results from a programme of work involving neural networks with inductive biases in the form of physical models. Firstly, motivated by the problem of learning from demonstration in robotics, we show how hybrid controllers can be learnt through switching density networks. Next, we show how variational recurrent neural network architectures can be used to perform data-efficient parameter estimation with dynamics observed in video streams. Finally, we show that these same methods can be applied to complex domains such as molecular geometry modelling. Using the same variational RNN architecture, we demonstrate efficient inversion of ultrafast X-ray scattering with dynamics constraints.