Source author record

Sungbin Lim

Explore this author’s publications, collaborators and research topics. This catalog record has not been claimed by its author.

ResearcherUnclaimed source record

math.AP Machine Learning Computer Vision Distributed, Parallel, and Cluster Computing math.FA math.PR

Catalog footprint

Research overview

9works

6topics

4close collaborators

Actions

Is this your work?

Open graph Browse papers

preprint2020arXiv

AutoCLINT: The Winning Method in AutoCV Challenge 2019

NeurIPS 2019 AutoDL challenge is a series of six automated machine learning competitions. Particularly, AutoCV challenges mainly focused on classification tasks on visual domain. In this paper, we introduce the winning method in the competition, AutoCLINT. The proposed method implements an autonomous training strategy, including efficient code optimization, and applies an automated data augmentation to achieve the fast adaptation of pretrained networks. We implement a light version of Fast AutoAugment to search for data augmentation policies efficiently for the arbitrarily given image domains. We also empirically analyze the components of the proposed method and provide ablation studies focusing on AutoCV datasets.

preprint2020arXiv

torchgpipe: On-the-fly Pipeline Parallelism for Training Giant Models

We design and implement a ready-to-use library in PyTorch for performing micro-batch pipeline parallelism with checkpointing proposed by GPipe (Huang et al., 2019). In particular, we develop a set of design components to enable pipeline-parallel gradient computation in PyTorch's define-by-run and eager execution environment. We show that each component is necessary to fully benefit from pipeline parallelism in such environment, and demonstrate the efficiency of the library by applying it to various network architectures including AmoebaNet-D and U-Net. Our library is available at https://github.com/kakaobrain/torchgpipe .

preprint2019arXiv

Fast AutoAugment

Data augmentation is an essential technique for improving generalization ability of deep learning models. Recently, AutoAugment has been proposed as an algorithm to automatically search for augmentation policies from a dataset and has significantly enhanced performances on many image recognition tasks. However, its search method requires thousands of GPU hours even for a relatively small dataset. In this paper, we propose an algorithm called Fast AutoAugment that finds effective augmentation policies via a more efficient search strategy based on density matching. In comparison to AutoAugment, the proposed algorithm speeds up the search time by orders of magnitude while achieves comparable performances on image recognition tasks with various models and datasets including CIFAR-10, CIFAR-100, SVHN, and ImageNet.

preprint2016arXiv

A Sobolev Space theory for stochastic partial differential equations with time-fractional derivatives

In this article we present an $L_p$ -theory ( $p\geq 2$ ) for the time-fractional quasi-linear stochastic partial differential equations (SPDEs) of type $\partial^α_tu=L(ω,t,x)u+f(u)+\partial^β_t \sum_{k=1}^{\infty}\int^t_0 ( Λ^k(ω,t,x)u+g^k(u))dw^k_t,$ where $α\in (0,2)$ , $β<α+\frac{1}{2}$ , and $\partial^α_t$ and $\partial^β_t$ denote the Caputo derivative of order $α$ and $β$ respectively. The processes $w^k_t$ , $k\in \mathbb{N}=\{1,2,\cdots\}$ , are independent one-dimensional Wiener processes defined on a probability space $Ω$ , $L$ is a second order operator of either divergence or non-divergence type, and $Λ^k$ are linear operators of order up to two. The coefficients of the equations depend on $ω(\in Ω), t,x$ and are allowed to be discontinuous. This class of SPDEs can be used to describe random effects on transport of particles in medium with thermal memory or particles subject to sticking and trapping.

preprint2015arXiv

An $L_q(L_p)$ -theory for parabolic pseudo-differential equations: Calderón-Zygmund approach

In this paper we present a Calderón-Zygmund approach for a large class of parabolic equations with pseudo-differential operators $\mathcal{A}(t)$ of arbitrary order $γ\in(0,\infty)$ . It is assumed that $\cA(t)$ is merely measurable with respect to the time variable. The unique solvability of the equation $$ \frac{\partial u}{\partial t}=\cA u-λu+f, \quad (t,x)\in \fR^{d+1} $$ and the $L_{q}(\fR,L_{p})$-estimate $$ \|u_{t}\|_{L_{q}(\fR,L_{p})}+\|(-Δ)^{γ/2}u\|_{L_{q}(\fR,L_{p})} +λ\|u\|_{L_{q}(\fR,L_{p})}\leq N\|f\|_{L_{q}(\fR,L_{p})} $$ are obtained for any $λ> 0$ and $p,q\in (1,\infty)$ .

preprint2015arXiv

An $L_q(L_p)$ -theory for the time fractional evolution equations with variable coefficients

We introduce an $L_q(L_p)$ -theory for the quasi-linear fractional equations of the type $\partial^α_t u(t,x)=a^{ij}(t,x)u_{x^i x^j}(t,x)+f(t,x,u), \quad t>0, \,x\in \mathbf{R}^d.$ Here, $α\in (0,2)$ , $p,q>1$ , and $\partial^α_t$ is the Caupto fractional derivative of order $α$ . Uniqueness, existence, and $L_q(L_p)$ -estimates of solutions are obtained. The leading coefficients $a^{ij}(t,x)$ are assumed to be piecewise continuous in $t$ and uniformly continuous in $x$ . In particular $a^{ij}(t,x)$ are allowed to be discontinuous with respect to the time variable. Our approach is based on classical tools in PDE theories such as the Marcinkiewicz interpolation theorem, the Calderon-Zygmund theorem, and perturbation arguments.

preprint2015arXiv

Asymptotic behaviors of fundamental solution and its derivatives related to space-time fractional differential equations

Let $p(t,x)$ be the fundamental solution to the problem $\partial_{t}^αu=-(-Δ)^βu, \quad α\in (0,2), \, β\in (0,\infty).$ In this paper we provide the asymptotic behaviors and sharp upper bounds of $p(t,x)$ and its space and time fractional derivatives $D_{x}^{n}(-Δ_x)^γD_{t}^σI_{t}^δp(t,x), \quad \forall\,\, n\in\mathbb{Z}_{+}, \,\, γ\in[0,β],\,\, σ, δ\in[0,\infty),$ where $D_{x}^n$ is a partial derivative of order $n$ with respect to $x$ , $(-Δ_x)^γ$ is a fractional Laplace operator and $D_{t}^σ$ and $I_{t}^δ$ are Riemann-Liouville fractional derivative and integral respectively.

preprint2015arXiv

Parabolic Littlewood-Paley inequality for a class of time-dependent operators of arbitrary order, and applications to higher order stochastic PDE

In this paper we prove a parabolic version of the Littlewood-Paley inequality for a class of time-dependent local and non-local operators of arbitrary order, and as an application we show this inequality gives a fundamental estimate for the $L_p$ -theory of the stochastic partial differential equations.

preprint2014arXiv

Parabolic BMO estimates for pseudo-differential operators of arbitrary order

In this article we prove the BMO- $L_{\infty}$ estimate $\|(-Δ)^{γ/2} u\|_{BMO(\mathbf{R}^{d+1})}\leq N \|\frac{\partial}{\partial t}u-A(t)u\|_{L_{\infty}(\mathbf{R}^{d+1})}, \quad \forall\, u\in C^{\infty}_c(\mathbf{R}^{d+1})$ for a wide class of pseudo-differential operators $A(t)$ of order $γ\in (0,\infty)$ . The coefficients of $A(t)$ are assumed to be merely measurable in time variable. As an application to the equation $\frac{\partial}{\partial t}u=A(t)u+f,\quad t\in \mathbf{R}$ we prove that for any $u\in C^{\infty}_c(\mathbf{R}^{d+1})$ $\|u_t\|_{L_p(\mathbf{R}^{d+1})}+\|(-Δ)^{γ/2}u\|_{L_p(\mathbf{R}^{d+1})}\leq N\|u_t-A(t)u\|_{L_p(\mathbf{R}^{d+1})},$ where $p\ in (1,\infty)$ and the constant $N$ is independent of $u$ .

Inspect adjacent papers, topics, institutions and collaborators without losing the researcher page.

Building this map preview

BZPEER is loading the nearby papers, people, topics and institutions for this page.

Sungbin Lim

Research overview

Is this your work?

9 publications

AutoCLINT: The Winning Method in AutoCV Challenge 2019

torchgpipe: On-the-fly Pipeline Parallelism for Training Giant Models

Fast AutoAugment

A Sobolev Space theory for stochastic partial differential equations with time-fractional derivatives

An $L_q(L_p)$ -theory for parabolic pseudo-differential equations: Calderón-Zygmund approach

An $L_q(L_p)$ -theory for the time fractional evolution equations with variable coefficients

Asymptotic behaviors of fundamental solution and its derivatives related to space-time fractional differential equations

Parabolic Littlewood-Paley inequality for a class of time-dependent operators of arbitrary order, and applications to higher order stochastic PDE

Parabolic BMO estimates for pseudo-differential operators of arbitrary order

See the researcher in context

Building this map preview

Sungbin Lim

Research overview

Is this your work?

9 publications

AutoCLINT: The Winning Method in AutoCV Challenge 2019

torchgpipe: On-the-fly Pipeline Parallelism for Training Giant Models

Fast AutoAugment

A Sobolev Space theory for stochastic partial differential equations with time-fractional derivatives

An Lq(Lp)L_q(L_p)Lq​(Lp​)-theory for parabolic pseudo-differential equations: Calderón-Zygmund approach

An Lq(Lp)L_q(L_p)Lq​(Lp​)-theory for the time fractional evolution equations with variable coefficients

Asymptotic behaviors of fundamental solution and its derivatives related to space-time fractional differential equations

Parabolic Littlewood-Paley inequality for a class of time-dependent operators of arbitrary order, and applications to higher order stochastic PDE

Parabolic BMO estimates for pseudo-differential operators of arbitrary order

See the researcher in context

Building this map preview

An $L_q(L_p)$ -theory for parabolic pseudo-differential equations: Calderón-Zygmund approach

An $L_q(L_p)$ -theory for the time fractional evolution equations with variable coefficients