6/23 (수)
 

09:30 - 11:00 
제목 Origin of hybrid percolation transitions: Jamming and self-organized criticality
연사 강병남 (서울대학교)
요약 Percolation is a geometric phase transition from a non-percolating state to a percolating state in the Euclidean space. This notion is not limited to the systems in Euclidean space but extends to complex networks, in which a giant cluster is regarded as a spanning cluster. This percolation transition (PT) has been widely used for understanding network resilience, community formation, the spread of disease in a population, and so on. A PT is generally considered to be one of the most robust continuous transitions; however, PTs in complex systems often exhibit various types such as continuous, discontinuous, hybrid, and infinite-order phase transitions. A hybrid PT (HPT) exhibits properties of the first-order and second-order transitions at the same transition point. One of the examples is k-core percolation in cluster pruning process. When a node is deleted, avalanche dynamics proceeds to keep the k core of a network, in which degree of each node is at least k. The number of nodes deleted during these consecutive pruning processes is considered as the avalanche size. This avalanche process is a critical branching process, and thus the dynamic critical exponents is related to the conventional static critical exponents of the PT. Another example occurs in the cluster merging process. The so-called restricted ER (r-ER) model exhibits such features. This r-ER model exhibits an HPT in which the order parameter jumps from zero to a finite value at a transition point, and the size distribution of the finite clusters exhibits power-law decay, indicating underlying critical behavior. In this model, clusters are partitioned into two sets; set A contains the smallest clusters, and set B contains the remaining large clusters. The rule of cluster merging is assigned asymmetric for sets A and B. Due to this asymmetry, clusters move back and forth between the two sets, and the intervals between the two crossing events, similar to the one observed in the tug-of- war play, exhibit a critical behavior, leading to a HPT. A question arises as to whether an HPT in cluster merging process occurs only in the system with such asymmetric rule and the tug-of-war process is fundamental. Here, we show that jamming and relaxation dynamics in drag-forced systems can be the origin of an HPT in cluster merging process.
 

11:00 - 12:00 
제목 TBA
연사 윤혜진  (Northwestern University)
요약
 

14:00 - 15:30 
제목 Viewing the generalization of overparameterized machine learning models through the lens of optimization
연사 신진우 (KAIST) 
요약 Classical results in the statistical learning theory suggest the following relationship between the model complexity and the generalization loss: If the model is too small, we cannot fit the dataset accurately, but the generalization loss will be small. If the model is too big, we can accurately fit the dataset, but the generalization loss will be large. However, modern empirical results on deep learning reveal that this picture is far from complete; overparameterized models often generalize better than smaller models. In this talk, I describe recent advances in the learning theory that attempts to solve this generalization mystery by understanding the implicit bias imposed by the deep learning optimization procedure. In particular, we discuss how the minimum-norm bias and sparse subnetwork bias can be used to explain the unexpectedly good generalization capabilities of overparameterized neural networks.
 

15:30 - 16:30 
제목 Explaining machine learning with a phase transition and vice versa
연사 김동희 (GIST)
요약 The talk comprises two parts. The first part addresses an open question of how the data-driven prediction works and how much one can trust such prediction by presenting our case study of finding a critical temperature of the phase transition in the Ising model via supervised learning. We have provided a minimal setting to show that the universality class of the training dataset limits the reusability of the trained machine for retrieving an exact critical temperature in similar systems. The second part presents our ongoing work of applying a neural quantum state to characterize the universality class of the long-range antiferromagnetic quantum Ising chain that has been under debate for some years. The variational Monte Carlo calculations with the restricted Boltzmann machine determine the critical exponents and central charge, strongly suggesting that the Ising university class is intact at any decay exponent of the long-range antiferromagnetic interaction.
 

16:30 - 17:30
제목 Jamming in multilayer supervised learning models and the force balance near the criticality
연사 황성민  (CFM)
요약 When a system of spheres is compressed, the spheres first move to empty space so as not to overlap with each other. As the volume decreases further, they reach a point where they can no longer move but start to form a rigid contact network. When such volume is reached, we say that the system undergoes a jamming transition.
The jamming transition of hard and soft spheres has been extensively studied in the last 20 years. Jammed packings of spheres display several critical properties. In particular, their pressure and force distributions are described by a set of power laws exponents independent of microscopic details of spheres. Recently the behaviour of the jamming transition has been exactly solved for spheres in the limit of infinite spatial dimensions. One outcome of these studies is the exact determination of the associated power-law exponents. Remarkably, numerical simulations show that these exponents appear to be independent of the dimensionality of the system and well fitted by their infinite-dimensional values, thus suggesting that universality is emerging at this transition.
An important step in this direction has been made by recognizing that jamming can be seen as a satisfiability transition for random constraint satisfaction problems with continuous degrees of freedom. Prototypes of such problems are classification problems in machine learning. It has been shown that the simplest of these models, the perceptron, is characterized by a jamming transition that is in the same universality class as hard and soft spheres.
In this work, we investigate further these perspectives. We consider large complex machine learning models, namely, multilayer supervised learning machines as continuous constraint satisfaction problems. We tackle these models analytically and we show that the resulting equations are a multi-dimensional extension of the previous single-field examples. Surprisingly we established an important dimensional reduction mechanism through which at jamming these models flow back to the hard spheres universality class. Finally, we discuss how these critical exponents determine the relaxation time scale of gradient descent dynamics using the force balance equation.

6/24 (목)

09:30 - 11:00
제목 Enjoy If Unavoidable: Genetic Noise from Control Perspective
연사 김철민  (UNIST)
요약 Origins and consequences of cell-to-cell variability are essential to the understanding of diverse biological processes underlying the development, aging, immune response, and tumorigenesis, just to name a few. As a micrometer-sized chemical reactor, living cells call for analytical frameworks that respect the stochasticity of biochemical reactions and discrete nature of macromolecules. In the earlier part of the talk, I propose some rationales of genetic noise for tuning the functional stability of a simple synthetic gene switch. The latter part of the talk will focus on the effects of zygosity of diploid cells, which adds another dimension to stochastic gene expression. I will introduce the diploid gene expression systems with homo- and heterozygous combination of alleles in the cis-regulatory sequences and characterize the noise profiles associated with zygosity. An emerging feat of heterozygosity is its counterintuitive capacity for genetic noise control, which offers a novel insight into the rich repertoire of balancing selection enriched in the regulatory sequences of the immune response genes.

11:00 - 12:00 
제목 Scaling behaviors of deep neural networks
연사 이재훈 (Google Research)
요약 For a large variety of models and datasets, neural network performance has been empirically observed to scale as a power-law with model size and dataset size. We would like to understand why these power laws emerge, and what features of the data and models determine the values of the power-law exponents. Since these exponents determine how quickly performance improves with more data and larger models, they are of great importance when considering whether to scale up existing models. In this talk, we’ll survey some of the well known power-law scaling behavior observed in deep learning. Drawing intuition from statistical physics, we observe that a simplifying limit arises as one scales up deep learning models. We’ll talk about a theoretical framework that explains and connects various scaling laws. In our simple framework, we identify variance-limited and resolution-limited scaling behavior for both dataset and model size, for a total of four scaling regimes.
 

14:00 - 15:30 
제목 Media effect on opinion formation
연사 김범준 (성균관대학교)
요약 Our opinions on a social issue can be affected by others’ opinions in social networks and also by various media we are acquainted with. In a modern society, there are many different media we can choose, and we often choose the one that is close to our own political and cultural tastes. We introduce a simple model in which the opinion of an agent is affected not only by other agents in the system, but also by the media. The effect by the media is tuned by a parameter, the media field F in our model, which can either strengthen (for F > 0) or weaken (for F < 0) the opinion of the agent. As F is varied, we find that our model exhibits three different states: neutral state, consensus state, and polarized state. We observe that a discontinuous transition occurs between the neutral and consensus states, and examine how the finiteness of the system size affects the transition between the consensus and polarized states. 
 

15:30 - 16:30 
제목 Message-passing theory of epidemics on complex networks
연사 민병준 (충북대학교)
요약 How to predict and control epidemics on a networked system is an important problem of much interest. Previous attempts in this field have mostly focused on the mean-field analysis, so that a full analysis of epidemic spreading at the level of a single node is lacking. In this work, we suggest a message-passing approach in order to fully analyze epidemic spreading on complex networks. In particular, we study i) mutually cooperative epidemics and ii) epidemics on a meta-population model, by using the message-passing equations. In addition, we suggest how to identify precisely influential spreaders in epidemics on a network.
 

16:30 - 17:30 
제목 Absorbing state phase transition in open quantum many-body systems
연사 조민재 (서울대학교)
요약 Recent advances in cold atomic physics offer a platform to explore non-equilibrium phase transition in open quantum many-body systems. Such non-equilibrium critical phenomena originate from the competition between quantum fluctuations (coherent Hamiltonian) and classical fluctuations (incoherent dissipation). For these systems, questions arise as to whether the competition between quantum coherent and classical incoherent fluctuations produces another type of universal behavior and the conditions under which they exhibit classical critical behavior in terms of the loss rates to the environment. In this talk, I aim to address these questions by considering the quantum contact process, which is a paradigmatic quantum non-equilibrium model with an absorbing state. I will discuss recent researches on this subject, ranging from the experimental realization of the model, the semi-classical mean-field approach, and the most recent low-dimensional quantum Monte Carlo simulations and a neural network approach.
 

6/25 (금)
 

09:30 - 11:00 
제목 Data Science for Social Impact
연사 차미영 (IBS/KAIST)
요약 Artificial intelligence (AI) and big data are bringing innovations to many research fields that have a direct social impact. In this talk, I’d like to share recent efforts on research related to Sustainable Development Goals (SDGs). One of them is the inference of economic activities based on satellite imagery. Recent advances in Computer Vision algorithms and the availability of high-resolution satellite images for remote sensing help us tackle this problem from a new perspective (i.e., AI + Geography + Economy). In particular, this talk will feature our group’s efforts in one of the most remote and closed areas in the world, North Korea. The talk will show how a human-machine collaborative algorithm can compute, for the first time, local-level and district-level estimates of economic activity from publicly available satellite imagery. The multi-faceted evaluation based on partial statistics confirms that our method, leveraging no label information, can explain up to 80% of the regional variation. Efforts like this on reliable and timely measurements of economic activities are fundamental for understanding economic development and designing government policies. 
 

11:00 - 12:00 
제목 Connectomics of Terabyte- to Petabyte-scale image data
연사 김진섭 (성균관대학교)
요약 To understand the fundamental principles of brain functions, identifying the neuronal types and mapping the connectivity is critical. High resolution electron microscope (EM) has been used to study the ultrastructure of neurons and synaptic connections. EM is the only available technology which enables the observation of the complete synaptic connectivity of entire neurons in a brain sample, the connectome. Terabytes to petabytes of brain images can be scanned thanks to the advancement in high-throughput serial EM imaging technologies. Computational methods to analyze those images are also attainable. In this presentation, I will introduce how we have employed these technologies to study the ultrastructures, principles of organizations, and computation mechanisms of various neural systems. As a perspective, I will argue that computational and connectomic analysis of EM image data is a way to study the fundamental principles of the brain.
Keywords:  neural networks, connectome, electron microscope, image analysis, ultrastructure, neural computation, principles of brain function 
 
14:00 - 15:00
제목 Scale-invariant representation of machine learning
연사 조정효 (서울대학교) 
요약 The success of machine learning relies on its structured representation of data. Similar data have close representations Z as compressed encoding for classification, or emergent labels for clustering. Universally in the supervised and unsupervised learning, we find that the frequency of Z follows Zipf's law with a few frequent and many rare representations. Using information theory, we derive how the Zipf's law can naturally arise without design in machine learning. The scale-invariant distribution corresponds to a maximally uncertain one among possible distributions of Z that guarantee pre-specified learning accuracy. Then we demonstrate that the characteristic representation has a functional advantage to largely compress frequent typical data, and clearly differentiate many atypical data as outliers. 
 

15:00 - 16:00 
제목 The complex journey for a sustainable power system
연사 김희태 (한국에너지공과대학교)
요약 A power grid is the biggest engineered system in the world, and it is evolving in response to the demand from society. To supply electricity mitigating the climate crisis, great ideas, including renewable energy technologies and smart grid, boost the challenge to green and quality electricity generation. The challenge makes the power system complex. Thus stable operation and management become another difficult problem to solve. In this talk, we overview the studies that theoretically investigate complex power systems and see the next target where we should go.
 

16:00 - 17:00
제목 Ecological communities from generalized Lotka-Volterra dynamics with random heterogeneous interactions
연사 박혜진 (APCTP)
요약 Natural populations ranging from microbial communities to animal societies consist of many different species. Some species compete to exploit a shared resource, whereas others can help each other. Such interactions affect the death or reproduction of species, thus shaping the composition of populations. The generalized Lotka-Volterra model has captured such large-scale interacting populations as a basic model for studying evolutionary ecologies. Combined with the random matrix theory, open questions about diversity--how can many different species coexist together? and how does the interaction types and diversity be related?--have been recently investigated. The results have shown that both interaction type and strength determine the diversity of the ecosystem and the correlation between interactions of survived species. However, the method is only applied to the homogeneous interaction structure while the interactions are generally heterogeneous in nature. Extending the previous approach, we investigate how heterogeneity of interactions plays a role to shape diversity. Our preliminary results show that heterogeneity can induce the catastrophe of species reducing the number of coexisting species.