학과 세미나 및 콜로퀴엄
In this talk, we discuss the Neural Tangent Kernel. The NTK is closely related to the dynamics of the neural network during training via the Gradient Flow(or Gradient Descent). But, since the NTK is random at initialization and varies during training, it is quite delicate to understand the dynamics of the neural network. In relation to this issue, we introduce an interesting result: in the infinite-width limit, the NTK converge to a deterministic kernel at initialization and remains constant during training. We provide a brief proof of the result for the simplest case.
Questions of parameter estimation – that is, finding the parameter values that allow a model to best fit some data – and parameter identifiability – that is, the uniqueness of such parameter values – are often considered in settings where experiments can be repeated to gain more certainty about the data. In this talk, however, I will consider parameter estimation and parameter identifiability in situations where data can only be collected from a single experiment or trajectory. Our motivation comes from medical settings, where data comes from a patient; such limitations in data also arise in finance, ecology, and climate, for example. In this setting, we can try to find the best parameters to fit our limited data. In this talk, I will introduce a novel, alternative goal, which we refer to as a qualitative inverse problem. The aim here is to analyze what information we can gain about a system from the available data even if we cannot estimate its parameter values precisely. I will discuss results that allow us to determine whether a given model has the ability to fit the data, whether its parameters are identifiable, the signs of model parameters, and/or the local dynamics around system fixed points, as well as how much measurement error can be tolerated without changing the conclusions of our analysis. I will consider various classes of model systems and will illustrate our latest results with the classic Lotka-Volterra system.
In the past decades, there has been considerable progress in the theory of random walks on groups acting on hyperbolic spaces. Despite the abundance of such groups, this theory is inherently not preserved under quasi-isometry. In this talk, I will present our study of random walks on groups that satisfy a certain QI-invariant property that does not refer to an action on hyperbolic spaces. Joint work with Kunal Chawla, Kasra Rafi, and Vivian He.
In this talk, we discuss some concepts that are used to study (hyperbolic) holomorphic dynamics on K3 surfaces. These concepts include Green currents, their laminations, and Green measures, which emerge as the natural choice for measuring maximal entropy.
These tools effectively establish Kummer rigidity – that is, when the Green measure is absolutely continuous to the volume measure, the surface is Kummer, and the dynamics is linear. We provide an overview of the techniques employed to establish this principle and provide a glimpse into their extension within the hyperkähler context – one of the higher-dimensional analogues of K3 surfaces.
Hénon maps were introduced by Michel Hénon as a simplified model of the Poincaré section of the Lorenz model. They are among the most studied discrete-time dynamical systems that exhibit chaotic behavior. Complex Hénon maps in any dimension have been extensively studied over the last three decades, in parallel with the development of pluripotential theory. We will present the dynamical properties of these maps such as the behavior of point orbits, variety orbits, equidistribution of periodic points and fine ergodic properties of the systems. This talk is based on the work of Bedford, Fornaess, Lyubich, Sibony, Smillie, and on recent work of the speaker in collaboration with Bianchi and Sibony.
Let V be a suvariety of a manifold M. We say that V has extension property,
if any bounded holomorphic function on V extends to a holomorphic function on M
with the same sup-norm. In the talk we shall explain connections between this
problem and operator theory (von Neumann inequality, interpolation problem)
as well as with the theory of invariant functions and metrics
We review constructions of Manolescu’s Floer homotopy type, which gives a homotopical refinement of monopole Floer homology. Based on it, we will introduce some homology cobordism/ knot concordance invariant. Using these invariants, we provide relative versions of 10/8 inequalities for 4-manifolds with boundary or surfaces in 4-manifolds. In particular, I’ll explain Manolescu’s relative 10/8 inequality, real 10/8 inequality, and Montague’s 10/8 inequality.
We first review fundamental concepts about Seiberg-Witten theory for closed 4-manifolds. Subsequently, we will introduce a refinement of Seiberg-Witten invariant, called Bauer—Furuta invariant. Using Bauer—Furuta invariant, I will explain how to prove Furuta’s 10/8 inequality and its variant for group actions proven by Bryan and Kato.