김세훈
카카오브레인Denoising diffusion-based generative models iteratively refines corrupted samples from white noises to the points in the original data space. This family of generative model can be easily trained by explicitly maximizing the data likelihood, thereby covering all modes in the data space, when the model capacity is sufficient. This now becomes the standard approach in many interesting text-to-image generation models, including DALL-E 2, Imagen, and Stable Diffusion. In this tutorial, I will first cover the basic idea of denoising diffusion models and detailed mathematical derivation to understand the underlying procedure of diffusion models. In specific, I will describe the main idea of DDPM, Improved DDPM, and DDIM. Then, I will discuss the technical challenges encountered in generating high-resolution images based on the text prompts, and how the well-known models (unCLIP, Imagen, Stable Diffusion) can resolve these challenges. Lastly, if the time allowed, I would describe kakaobrain’s approaches to contribute the research community in this field by introducing COYO datasets (700M image-text pairs) and Karlo (text-conditional image generation model).
김동우
POSTECHNormalizing flows (NFs) offer an answer to a long-standing question in machine learning: How one can define faithful probabilistic models for complex high-dimensional data. NFs solve this problem by means of non-linear bijective mappings from simple distributions (e.g. multivariate normal) to the desired target distributions. These mappings are implemented with invertible neural networks and thus have high expressive power and can be trained by gradient descent in the usual way. Compared to other generative models, the main advantage of normalizing flows is that they can offer exact and efficient likelihood computation and data generation thanks to bijectivity. This session will explain the theoretical underpinnings of NFs, show various practical implementation options, clarify their relationships with the other generative models such as GANs and VAEs.
신현정
아주대학교In classical kernel methods, kernel embedding can be viewed as a generalization of the individual data-point feature mapping as done in support vector machines and other kernel methods. Recently, it has meanwhile been extended to the embedding of distributions into infinite-dimensional feature spaces that can preserve all of the statistical features of arbitrary distributions, while allowing one to compare and manipulate distributions using Hilbert space operations such as inner products, distances, projections, linear transformations, and spectral analysis. We begin with a brief introduction to a reproducing kernel Hilbert space (RKHS) in which the key arsenal of kernel methods. Then positive definite kernels which forms the backbone of this research area are explained followed by introduction to various kernels proposed for various data types which are vectors, strings, graphs, images, time series, sequences, manifolds, and other structured objects. Then we gear up to simple embedding techniques such as kPCA, kCCA, kICA, etc. Next, we discuss the embedding of distributions which enables us to apply RKHS methods a wide range of applications such as sample testing and domain adaptation. Lastly, we discuss relationships between kernel embedding and other related areas.
최승진
IntellicodeRandom projection is a probabilistic dimensionality reduction method where a high-dimensional data point is projected onto a column space of a random matrix to construct a low-dimensional representation. What is good about this is that the projection matrix does not have to be learned from data, unlike most of dimensionality reduction methods such as principal component analysis or Fisher’s discriminant analysis. In other words, random projection is a data-independent method. What is even awesome characteristics of random projection is that it guarantees an epsilon-distortion distance preserving embedding into a lower-dimensional space, which is well justified by the celebrating Johnson-Lindenstrauss (JL) lemma. It has been extensively studied in machine learning and theoretical computer science. In the first half of this tutorial, I will begin with the linear projection and will review the random projection and JL lemma. The second half of the talk will be on nonlinear random projection. Of particular interests are random Fourier features which scales up kernel machines. A successful example is the ‘random kitchen sinks’ which scales up the kernel regression as well as deep neural networks, via randomized feature map instead of optimizing feature representations. If time allowed, I will also mention a method of speeding up random kitchen sinks, which is known as ‘fastfood’.
노영균
한양대학교/고등과학원In this lecture, we review how we can take advantage of using information theory. As a theory, it provides referential and comprehensive concepts for practical applications. I will explain the theoretical motivations and justifications for many machine learning algorithms. The explanations will include various methods for feature selection, regularization and in particular, recent advances in the estimation of information contents for various applications.
김선
서울대학교이상선
서울대학교In this talk, we survey recent developments in graph-level learning. Traditional and widely investigated topics on graph learning are mostly on node-level and edge-/link prediction in networks, e.g., in social networks. Graph learning has become more important in scientific and medical domains, which is the topic of this tutorial lecture. Graph learning in scientific and medical domains is fundamentally different from traditional graph learning in two respects. First, in these domains, an individual, e.g., a patient and a chemical compound, is a graph where an individual is a whole graph and nodes in each graph are features of an individual such as atoms or genes. Note that in social networks, individuals are nodes. Second, graph learning in these domains is to discover distinguishing features of patient groups or chemical compound sets and classify them in terms of annotated characteristics such as cancer metastasis or toxicity, thus mining on a number of graphs. This tutorial is to survey recent developments in this topic. There are a number of research questions in this topic. First, graph construction is not straightforward and construction of graphs from data requires thoughtful strategies. Second, there are usually small number of samples in these domains, thus graph augmentations is another important research question. Third, in the medical domains, graph learning needs to require decoding of complex interactions of features, e.g., genes. Fourth, in the chemical domain, graphs vary significantly in terms of graph size, and graph mining in the chemical domain requires to optimize quite a number of tasks simultaneously, thus multi-task learning with graphs of varying sizes. Lastly, in the drug repositioning task, it is necessary to deal with heterogeneous networks where nodes are of completely different entities such as genes, diseases, and drug compounds. Mining associations in these heterogeneous networks require sophisticated computational strategies. In this tutorial, we will define research questions and survey recent developments in each of research questions with limitations of current technologies and future directions.
김건희
서울대학교본 강연에서는 대규모 모델의 학습의 기반이 되는 자기감독학습 (Self-supervised learning)의 기초와 응용에 대해 살펴본다. 우선 자기감독학습의 폭발적 성장을 이끈 초기의 기본 모델에 대해 다루고, 다형식(Multimodal data) 데이터 기반 모델과 상업적으로 큰 성공을 거둔 방법론에 대해 논의한다. 마지막으로 현재의 연구 경향과 미래의 발전 방향에 대해 함께 생각해본다.
이근배
POSTECH본 발표는 최근의 자연어처리 연구에 있어서 심층신경망을 이용한 여러 연구접근에 대해서 설명하고자 한다. 도입에서는 자연어처리의 여러 응용분야 및 다루어야 할 단위 task 즉 문장분류, 품사태깅, 개체명태깅, 구문분석, 동시참조처리, 자동요약등의 문제에 대해 설명하고 왜 이러한 문제들이 기술적으로 어려운지 애매성 및 여러 예시를 들어 설명한다. 이후 이러한 task를 처리하기 위한 심층학습의 기본원리와 심층기계학습을 이용한 시퀀스처리에 대해 설명한다. 이는 향후 자연어에 있어 단어열을 처리하는 기본 원리가 된다. 자연어처리에서 가장 다루기 힘든 어휘 단계의 애매성을 처리하기 위한 워드임베딩 기본 원리를 설명하고 이를 확장한 문맥 기반 워드 임베딩에 대해 기본 원리와 응용을 설명한다. 이러한 문맥 기반 워드 임베딩을 만들기 위한 가장 각광을 받는 모델로 트랜스포머 구조와 응용을 설명하고 이를 이용한 버트 및 GPT 사전학습 구조를 다룬다. 이후 기계번역등 자연어처리의 여러 시퀀스태깅의 기본 문제 및 대표적 응용들인 자연어 기계독해와 자연어 챗봇 대화 시스템에 대해 하나의 심층신경망 모델로 전체 학습 가능한 모델에 대해 설명하고 본 발표를 요약하고자 한다.
Early Registration (~ 1월 8일) |
Late Registration (1월 9일 ~ 12일) |
||
---|---|---|---|
Academy | 교수 | 20만원 | 25만원 |
학생 | 10만원 | 15만원 | |
Industry | 20만원 | 25만원 |