Uploaded July 2022 | Updated November 2026, 3 days ago
Deep Equilibrium Models (DEQs) are a class of machine learning models that computes the fixed point of a single nonlinear operator in lieu of a traditional multi-layer network. Our recently proposed monotone equilibrium model (monDEQ) provides a firmer grounding for such models, by providing a parameterization that guarantees the existence and uniqueness of this fixed point. In this talk, I will highlight some of our recent work on these monDEQ models. Specifically, I will first highlight the general monDEQ framework, how it provides guarantees on fixed point uniqueness, and how it allows for efficient operator splitting methods for computing the fixed point; I will then illustrate how it enables us bound Lipchitz constant of the monDEQ models, which in turn allows us to produce generalization bounds for these architectures (non-trivial for these effectively "infinitely deep" networks); I will highlight a few different approaches to characterizing robustness of monDEQ models; and I will finish by highlighting the connections between monDEQs and mean-field inference, to illustrate how we can use these approaches to formulate Boltzman-machine-like models for which mean-field inference is guaranteed to converge to a globally optimal solution.
Slides: mathunion.org/fileadmin/IMU/ICM2022/Presentation-slides/183-Zico%20Kolter.pdf
Deep Equilibrium Models (DEQs) are a class of machine learning models that computes the fixed point of a single nonlinear operator in lieu of a traditional multi-layer network. Our recently proposed monotone equilibrium model (monDEQ) provides a firmer grounding for such models, by providing a parameterization that guarantees the existence and uniqueness of this fixed point. In this talk, I will highlight some of our recent work on these monDEQ models. Specifically, I will first highlight the general monDEQ framework, how it provides guarantees on fixed point uniqueness, and how it allows for efficient operator splitting methods for computing the fixed point; I will then illustrate how it enables us bound Lipchitz constant of the monDEQ models, which in turn allows us to produce generalization bounds for these architectures (non-trivial for these effectively "infinitely deep" networks); I will highlight a few different approaches to characterizing robustness of monDEQ models; and I will finish by highlighting the connections between monDEQs and mean-field inference, to illustrate how we can use these approaches to formulate Boltzman-machine-like models for which mean-field inference is guaranteed to converge to a globally optimal solution.
Slides: mathunion.org/fileadmin/IMU/ICM2022/Presentation-slides/183-Zico%20Kolter.pdf







![Zhifei Zhang: Hydrodynamic stability at high Reynolds number and Transition threshold problem
The hydrodynamic stability theory is mainly concerned with how the laminar flows become unstable and transit to turbulence at high Reynolds number. To shed some light on the transition mechanism, Trefethen et al. [Science 261(1993)] proposed the transition threshold problem. Many physical effects such as 3-D lift-up, inviscid damping, enhanced dissipation and boundary layer, play a crucial role in determining the transition threshold. In this talk, I will survey some important progress on linear inviscid damping and enhanced dissipation for shear flows. I will outline key ideas and main ingredients in our proof of transition threshold for 3-D Couette flow in a finite channel.
Slides: https://www.mathunion.org/fileadmin/IMU/ICM2022/Presentation-slides/79-Zhifei%20Zhang.pdf Zhifei Zhang: Hydrodynamic stability at high Reynolds number and Transition threshold problem](https://i.ytimg.com/vi/yyyjVT65s20/mqdefault.jpg)


