Virtual

Hilbert-Schmidt independence criterion (HSIC) is among the most widely-used approaches in machine learning and statistics to measure the independence of random variables. Despite its popularity and success in numerous applications, quite little is known about when HSIC characterizes independence. I am going to provide a complete answer to this question, with conditions which are often easy to verify in practice.

This talk is based on joint work with Bharath Sriperumbudur.

The blow-up B of a scheme X in a closed subscheme Z enjoys the universal property that for any scheme X' over X such that the pullback of Z to X' is an effective Cartier divisor, there is a unique morphism of X' into B over X. It is well-known that the blow-up commutes along flat base change.

In this talk, I will discuss a derived enhancement B' of B, namely the derived blow-up, which enjoys a universal property against all schemes over X, satisfies arbitrary (derived) base-change, and contains B as a closed subscheme. To this end, we will need some elements from derived algebraic geometry, which I will review along the way. This will allow us to construct the derived blow-up as the projective spectrum of the derived Rees algebra, and state its functor of points in terms of virtual Cartier divisors, using Weil restrictions.

This is based on ongoing joint work with Adeel Khan and David Rydh.

The on-shell superspace formulation of N=4 SYM allows the writing of all possible scattering processes in one compact object called the super-amplitude.   Famously, the super-amplitude integrand can be extracted from generalized polyhedra called the amplituhedron. In this talk, I will review this construction and present a natural generalization of the amplituhedron that we proved at tree level and conjectured at loop level to correspond to the product of two parity conjugate superamplitudes. The sum of all parity conjugate amplitudes corresponds to a particular limit of the supercorrelator through the Wilson Loop/Amplitude duality.  I will conclude by discussing this connection from a geometrical point of view. This talk is based on the reference arXiv:2106.09372 .

In this talk, I will introduce our investigations on how to make deep learning easier to optimize, faster to train, and more robust to out-of-distribution prediction. To be specific, we design a group-invariant optimization framework for ReLU neural networks; we compensate the gradient delay in asynchronized distributed training; and we improve the out-of-distribution prediction by incorporating “causal” invariance.

The introduction to the talk will review basics of $G_{2}$ geometry in seven dimensions, and associative and co-associative submanifolds. In one part of the talk we will explain how fibrations with co-associative fibres, near the “adiabatic limit” when the fibres are very small,  give insights into various questions about moduli spaces of $G_{2}$ structures and singularity formation. This part is mostly speculative. In the other part of the talk we discuss some analysis questions which enter when setting up the foundations of this adiabatic theory. These can be seen as codimension 2 analogues of free boundary problems and related questions have arisen in a number of areas of differential geometry recently.

I will discuss a string theory way to study G_2 instanton moduli space and explain how to compute the instanton partition function for a certain G_2 manifold. An important insight comes from the twisted M-theory on the G_2 manifold. Building on the example, I will explain a possibility to extend the story to a large set of conjectural G_2 manifolds and a possible connection to 4d N=1 SCFT via geometric engineering. This talk is based on https://arxiv.org/abs/2109.01110 and a series of works in progress with Michele Del Zotto and Yehao Zhou.

I'll describe an interacting holomorphic-topological field theory in eleven dimensions defined on products of Calabi-Yau 5-folds with real one-manifolds. The theory describes a certain deformation of the cotangent bundle to the moduli of Calabi-Yau deformations of the 5-fold and conjecturally describes a certain protected sector of eleven-dimensional supergravity. Strikingly, the theory has an infinite dimensional global symmetry algebra given by an extension of the exceptional lie superalgebra E(5,10) first studied by Kac. This talk is based on joint work with Ingmar Saberi and Brian Williams.

We discuss linear convolutional neural networks (LCNs) and their critical points. We observe that the function space (that is, the set of functions represented by LCNs) can be identified with polynomials that admit certain factorizations, and we use this perspective to describe the impact of the network's architecture on the geometry of the function space.

For instance, for LCNs with one-dimensional convolutions having stride one and arbitrary filter sizes, we provide a full description of the boundary of the function space. We further study the optimization of an objective function over such LCNs: We characterize the relations between critical points in function space and in parameter space and show that there do exist spurious critical points. We compute an upper bound on the number of critical points in function space using Euclidean distance degrees and describe dynamical invariants for gradient descent.

This talk is based on joint work with Thomas Merkh, Guido Montúfar, and Matthew Trager.

In embryonic development, formation of the retinal vasculature is  critically dependent on prior establishment of a mesh of astrocytes.  
Astrocytes emerge from the optic nerve head and then migrate over the retinal surface in a radially symmetric manner and mature through 
differentiation.  We develop a PDE model describing the migration and  differentiation of astrocytes, and numerical simulations are compared to 
experimental data to assist in elucidating the mechanisms responsible for the distribution of astrocytes via parameter analysis. This is joint 
work with Timothy Secomb.

In this talk, I will explain how to obtain the space of local operators of a 4d N=2 gauge theory using the category of line operators in the Kapustin twist (holomorphic topological twist). This category is given a precise definition by Cautis-Williams, as the category of equivariant coherent sheaves on the space of Braverman-Finkelberg-Nakajima. We compute the derived endomorphism of the monoidal unit in this category, and show that it coincides with the vacuum module of the Poisson vertex algebra of Oh-Yagi and Butson. The Euler character of this space reproduces the Schur index. I will also explain how to obtain the space of local operators at the junction of minimal Wilson-t’Hooft line operators. Its Euler character can be compared to the index formula of Cordova-Gaiotto-Shao. This is based on arXiv: 2112.12164.