SearcharxivSearch

arXiv subjects

Ying Hong Tham

Publications and source records attributed to Ying Hong Tham.

11 recordsLinked to original sources

LAMP: Look-Ahead Mixed-Precision Inference of Large Language Models

Mixed-precision computations are a hallmark of the current stage of AI, driving the progress in large language models towards efficient, locally deployable solutions. This article addresses the floating-point computation of compositionally-rich functions, concentrating on transformer inference. Based on the rounding error analysis of a composition $f(g(\mathrm{x}))$, we provide an adaptive strategy that selects a small subset of components of $g(\mathrm{x})$ to be computed more accurately while all other computations can be carried out with lower accuracy. We then explain how this strategy can be applied to different compositions within a transformer and illustrate its overall effect on transformer inference. We study the effectiveness of this algorithm numerically on GPT-2 models and demonstrate that already very low recomputation rates allow for improvements of up to two orders of magnitude in accuracy.

cs.LG

Generalization of the Thistlethwaite--Tsvietkova Method

Thurston's equations determine the hyperbolic structure of a 3-manifold with a triangulation. In work by Thistlethwaite and Tsvietkova, an alternative method was developed for link complements in $S^3$ depending on the link diagram, where a set of labels are associated to the vertices and edges of the link diagram, and one attempts to solve a set of equations on the labels. Under certain conditions, there exists a solution to these equations that corresponds to the complete hyperbolic structure, but in general it is difficult to determine which one it is. We generalize this method to 3-manifolds with a polyhedral decomposition, and show that solutions to the equations correspond to $PSL(2,\mathbb{C})$-representations of the fundamental group, and that the solution with the largest volume corresponds to the complete hyperbolic structure. We also consider different classes of complements of links, in particular links in the thickened torus and fully augmented links. For the latter, we establish a correspondence between solutions satisfying some criteria and circle packings realizing the region graph associated to the fully augmented link.

math.GT

On the Volume Density Spectrum of Fully Augmented Links

For a hyperbolic fully augmented link in $S^3$, its \emph{FAL volume density} is the ratio of its volume to the number of augmentations. We show that the set of FAL volume densities is dense in $[2\voct, 10\vtet)$, but discrete in $[\voct,2\voct)$.

math.GT

Excision for Spaces of Admissible Skeins

The skein module for a d-dimensional manifold is a vector space spanned by embedded framed graphs decorated by a category A with suitable extra structure depending on the dimension d, modulo local relations which hold inside d-balls. For a full subcategory S of A, an S-admissible skein module is defined analogously, except that local relations for a given ball may only be applied if outside the ball at least one edge is coloured in S. In this paper we prove that admissible skein modules in any dimension satisfy excision, namely that the skein module of a glued manifold is expressed as a coend over boundary values on the boundary components glued together. We furthermore relate skein modules for different choices of S, apply our result to cylinder categories, and recover the relation to modified traces.

math.QA

The Y-Product

We present a topological construction that provides many examples of non-commutative Frobenius algebras that generalizes the well-known pair-of-pants. When applied to the solid torus, in conjunction with Crane-Yetter theory, we provide a topological proof of the Verlinde formula. We also apply the construction to a solid handlebody of higher genus, leading to a generalization of the Verlinde formula (not the higher genus Verlinde formula); in particular, we define a generalized $S$-matrix. Finally, we discuss the relation between our construction and Yetter's construction of a handle as a Hopf algebra, and give a generalization.

math.QA

Hyperbolicity of Augmented Links in the Thickened Torus

For a hyperbolic link K in the thickened torus with no bigons, we show that there is a decomposition of the complement of a link L, obtained from augmenting K, into torihedra. We further decompose the torihedra into angled pyramids and finally angled tetrahedra. These fit into an angled structure on a triangulation of the link complement, and thus by [5], this shows that L is hyperbolic.

math.GT

Reduced Tensor Product on the Drinfeld Center

The annulus comes with a "stacking" operation which glues two annuli into one. This provides a tensor product structure on the category of boundary values $Z_{\text{CY}}(\text{Ann})$ associated to the annulus in an extended Crane-Yetter TQFT. It is known that $Z_{\text{CY}}(\text{Ann}) \simeq \mathcal{Z}(\mathcal{A})$, the Drinfeld center of the premodular category $\mathcal{A}$ from which $Z_{\text{CY}}$ is constructed. We give an explicit formula for the tensor product on $\mathcal{Z}(\mathcal{A})$ that corresponds to the stacking operation on $Z_{\text{CY}}(\text{Ann})$.

math.QA

Explicit Factorization of a Categorical Center

Given a braided fusion category $C$, it is well known that the natural map $C \boxtimes C^{bop} \to Z(C)$ from the square of $C$ to the (Drinfeld) categorical center $Z(C)$ is an equivalence if and only if $C$ is modular. However, it is not clear how to construct the inverse and the natural isomorphisms. In this work, we provide an explicit construction using insights from a specific quantum field theory, and explore how the equivalence fails for the degenerate cases.

math.QA

On the Category of Boundary Values in the Extended Crane-Yetter TQFT

The Crane-Yetter state sum is an invariant of closed 4-manifolds, defined in terms of a triangulation, based on 15-j symbols associated to the category A of representations over quantum sl2 (at a root of unity). In this thesis, we define the state sum in terms of a 'PLCW decomposition', which generalizes triangulations, and generalize A to an arbitrary premodular category. We extend the state sum to 4-manifolds with corners, making it an extended TQFT. We also develop a parallel theory based on skeins, which are essentially A-colored graphs, and we show that the two theories are equivalent. Focusing on the 2-dimensional part, we prove several properties of skein categories, the most important of which is that they satisfy excision. We provide explicit algebraic descriptions of the category associated to the once-punctured torus and the annulus, giving rise to a new tensor product on the Drinfeld center of a premodular category. As it is well-known that, when A is modular, the Crane-Yetter state sum computes the signature of a closed 4-manifold, we connect the Crane-Yetter theory to the signature of a 4-manifold with boundary and even corners. Finally, we show that the Reshetikhin-Turaev TQFT is a boundary theory of the Crane-Yetter theory (up to a normalization).

math.QA

Factorization Homology and 4D TQFT

In [BK], it is shown that the Turaev-Viro invariants defined for a spherical fusion category $\mathcal{A}$ extends to invariants of 3-manifolds with corners. In [Kir], an equivalent formulation for the 2-1 part of the theory (2-manifolds with boundary) is described using the space of "stringnets with boundary conditions" as the vector spaces associated to 2-manifolds with boundary. Here we construct a similar theory for the 3-2 part of the 4-3-2 theory in [CY1993].

math.QA

The Elliptic Drinfeld Center of a Premodular Category

Given a tensor category C, one constructs its Drinfeld center Z(C) which is a braided tensor category, having as objects pairs (X, lambda), where X in Obj(C) and lambda is a half-braiding. For a premodular category C, we construct a new category Zel(C) which we call the Elliptic Drinfeld Center, which has objects (X, lambda1, lambda2), where the lambda i's are half-braidings that satisfy some compatibility conditions. We discuss an SL2(Z)-action on Zel(C) that is related to the anomaly appearing in Reshetikhin-Turaev theory. This construction is motivated from the study of the extended Crane-Yetter TQFT, in particular the category associated to the once punctured torus.

math.QA