Liang Xu | EPFL Graph Search

Liang Xu is a postdoctoral scholar at École polytechnique fédérale de Lausanne, Switzerland. He received his Ph.D. degree in Electrical and Electronic Engineering from Nanyang Technological University, Singapore, in 2018 and the M.Sc. and B.Sc. degrees in Automation from Harbin Institute of Technology, China, in 2013 and 2011, respectively. Prior to his current position, he was a postdoctoral scholar at Nanyang Technological University, Singapore from 2018 to 2019. His research interests include reinforcement learning and networked control systems.

Hamiltonian Deep Neural Networks Guaranteeing Non-Vanishing Gradients by Design

Giancarlo Ferrari Trecate, Luca Furieri, Clara Lucía Galimberti, Liang Xu

Deep Neural Networks (DNNs) training can be difficult due to vanishing and exploding gradients during weight optimization through backpropagation. To address this problem, we propose a general class of Hamiltonian DNNs (H-DNNs) that stem from the discretization of continuous-time Hamiltonian systems and include several existing DNN architectures based on ordinary differential equations. Our main result is that a broad set of H-DNNs ensures non-vanishing gradients by design for an arbitrary network depth. This is obtained by proving that, using a semi-implicit Euler discretization scheme, the backward sensitivity matrices involved in gradient computations are symplectic. We also provide an upper-bound to the magnitude of sensitivity matrices and show that exploding gradients can be controlled through regularization. The good performance of H-DNNs is demonstrated on benchmark classification problems, including image classification with the MNIST dataset.

2023

Neural Energy Casimir Control for Port-Hamiltonian Systems

Giancarlo Ferrari Trecate, Liang Xu, Muhammad Zakwan

The energy Casimir method is an effective controller design approach to stabilize port-Hamiltonian systems at a desired equilibrium. However, its application relies on the availability of suitable Casimir and Lyapunov functions, whose computation are generally intractable. In this paper, we propose a neural network-based framework to learn these functions. We show how to achieve equilibrium assignment by adding suitable regularization terms in the training cost. We also propose a parameterization of Casimir functions for reducing the training complexity. Moreover, the distance between the equilibrium of the learned Lyapunov function and the desired equilibrium is analyzed, which indicates that for small suboptimality gaps, the distance decreases linearly with respect to the training loss. Our methods are backed up by simulations on a pendulum system.

2022

A unified framework for Hamiltonian deep neural networks

Giancarlo Ferrari Trecate, Clara Lucía Galimberti, Liang Xu

Training deep neural networks (DNNs) can be difficult due to the occurrence of vanishing/explod- ing gradients during weight optimization. To avoid this problem, we propose a class of DNNs stemming from the time discretization of Hamiltonian systems. The time-invariant version of the corresponding Hamiltonian models enjoys marginal stability, a property that, as shown in previ- ous works and for specific DNNs architectures, can mitigate convergence to zero or divergence of gradients. In the present paper, we formally study this feature by deriving and analysing the backward gradient dynamics in continuous time. The proposed Hamiltonian framework, besides encompassing existing networks inspired by marginally stable ODEs, allows one to derive new and more expressive architectures. The good performance of the novel DNNs is demonstrated on benchmark classification problems, including digit recognition using the MNIST dataset.

2021