유튜브 스트리밍 예정입니다: https://www.youtube.com/live/2vbdfFuT064

1. 일시: 2024년 11월 14일(목), 14:00-16:00

2. 장소: 판교 테크노밸리 산업수학혁신센터 세미나실

   • 경기 성남시 수정구 대왕판교로 815, 기업지원허브 231호 국가수리과학연구소
   • 무료주차는 2시간 지원됩니다.

3. 발표자: 최재성 박사(고등과학원)

4. 주요내용: How reservoir computing differs from conventional neural networks and its potential applications

Reservoir computing emerged as a distinctive paradigm in handling temporal tasks. This framework significantly reduces trainable parameters compared to conventional neural networks while maintaining computational performance. Beyond its initial appeal for efficiency, it has gained attention for its physical implementation and unique applications in nonlinear dynamics-based computing. This presentation explores how reservoir computing opens new possibilities in specialized tasks, including associative memory for dynamics and unsupervised denoising, demonstrating its complementary role to conventional neural networks.

유튜브 스트리밍 예정입니다.

1. 일시: 2024년 10월 30일(수), 14:00-16:00

2. 장소: 판교 테크노밸리 산업수학혁신센터 세미나실

   • 경기 성남시 수정구 대왕판교로 815, 기업지원허브 231호 국가수리과학연구소
   • 무료주차는 2시간 지원됩니다.

3. 발표자: 신선영 교수(포항공대)

4. 주요내용: Clustered hidden Markov models

We consider cases where a hidden Markov model (HMM) has so many distinct states that its state space needs a simpler structure. A novel framework we develop, named the clustered hidden Markov model (CHMM), decomposes the state space into sets of clustered states, by adding a hidden process to the HMM. The CHMM consists of two hidden processes and an observable process, where the first hidden process is a Markov process with clustered state representations and the second hidden process has the original state space. The second hidden process is second-order dependent on the first hidden process; that is, the current state of the second hidden process is dependent on the current and previous states of the first hidden process. The CHMM is a new approach to designing the structure of the hidden states while maintaining the information in the HMM. We derive two hidden Markov submodels from the CHMM. The first hidden process and the observable process of the CHMM form a hidden Markov model with additional dependencies while the second hidden process and the observable process of the CHMM form a classical hidden Markov model. A novel finding is that the hidden states corresponding to a clustered state share the same transition probabilities, which are represented as the identical transition matrix rows of the second hidden Markov model. The identical row structure of the transition matrix motivates us to consider the penalization method for learning the CHMM, simultaneously recovering the clustered state space. The penalized estimation maximizes the likelihood regularized by group smooth clipped absolute deviation (SCAD) penalty to all pairwise differences of the transition matrix rows. We establish the asymptotic properties of the penalized estimator. Simulation studies support the outperformance of our proposed method. Its application to protein structure sequence data demonstrates that the CHMM simplifies the classification of protein segments, and consequently improves interpretability.

유튜브 스트리밍 예정입니다.

1. 일시: 2024년 10월 29일(화), 10:30-12:30

2. 장소: 판교 테크노밸리 산업수학혁신센터 세미나실

   • 경기 성남시 수정구 대왕판교로 815, 기업지원허브 231호 국가수리과학연구소
   • 무료주차는 2시간 지원됩니다.

3. 발표자: 김혜현 교수(경희대학교)

4. 주요내용: Partitioned neural networks for partial differential equations

In this talk, I will give a brief introduction to neural network approximation to partial differential equations (PDEs). To deal with error and performance issues in the neural network approximation, partitioned neural networks (PNNs) can be used as a solution surrogate. PNNs are formed based on the partition of the problem domain and each network in PNNs approximates the PDE solution in each small subdomain. The use of PNNs allows more flexible local network and hyperparameter settings, and more efficient parameter training by parallel computation. In the second part of my talk, I will present iterative algorithms for PNNs, that can utilize the partitioned network structure for the parallel computation. Numerical results will be presented for test examples to show the performance and scalability of the proposed iterative algorithms.

1. 일시: 2024년 8월 29일(목), 14:30-17:30

2. 장소: 판교 테크노밸리 산업수학혁신센터 세미나실

   • 경기 성남시 수정구 대왕판교로 815, 기업지원허브 231호 국가수리과학연구소
   • 무료주차는 2시간 지원됩니다.

3. 14:30-16:00

   • 발표자: 서효원 교수(국립군산대학교)
   • 주제: Multi-agent system on a sphere
   • 초록: As technology advances, multi-agent systems operating throughout the entire Earth’s surface have become feasible. We are concerned with a multi-agent system, which contains centripetal force, inter-particle bonding force, and velocity-alignment operator. In particular, we employ a regularized rotation operator to overcome the singularity problem at antipodal points. In this talk, we introduce several control laws for given tasks, such as target tracking or formation. Using these control laws, we observe the asymptotic behavior of agents: complete or practical synchronization. Our numerical simulations show that it is applicable to reality. The presented results come from a series of work with S.-H. Choi and D. Kwon.

4. 16:00 – 17:30

   • 발표자: 최선호 교수(경희대학교)
   • 주제: MULTI-AGENT SYSTEMS USING BEARING ANGLES AND ANCHOR POINTS
   • 초록: We propose a bearing-based control law for double-integrator multi-agent systems. The objective is that each agent converges towards a predefined point. Since bearing angles alone cannot determine a unique formation, at least three anchor points are required. The leader agents are those converging towards these anchor points. These leader agents have access to the location information of the anchor points, their own location information, and the bearing angles of surrounding agents. We assume that all other agents can only collect bearing angle information from neighboring agents. Like most bearing-based models, our model includes singularities, preventing the use of LaSalle’s invariance principle. To overcome this, we introduce a new energy functional. Using this energy functional and Barbalat’s lemma, we prove that all global solutions of this equation converge to the specified formation.

현장 강의만 진행될 예정입니다.

1. 일시: 2024년 8월 26일(월), 14:00-16:00

2. 장소: 판교 테크노밸리 산업수학혁신센터 세미나실

   • 경기 성남시 수정구 대왕판교로 815, 기업지원허브 231호 국가수리과학연구소
   • 무료주차는 2시간 지원됩니다.

3. 발표자: 김미란 교수(한양대학교)

4. 주요내용: Homomorphic encryption

Homomorphic encryption is an encryption method to allow arbitrary computation on encrypted data without decryption. It has emerged as one of the promising solutions to address privacy and security issues in outsourcing computation on sensitive data. In this talk, I will introduce recent development and progress on homomorphic encryption. In particular, I will introduce a homomorphic encryption cryptosystem that enables approximate computation over encryption, which has shown remarkable performance in artificial intelligence. I will then summarize the state-of-art benchmarks of the encryption systems in real-world applications such as machine learning and biomedical study.

유튜브 스트리밍 예정입니다.

1. 일시: 2024년 8월 21일(수), 10:30-12:30

2. 장소: 판교 테크노밸리 산업수학혁신센터 세미나실

   • 경기 성남시 수정구 대왕판교로 815, 기업지원허브 231호 국가수리과학연구소
   • 무료주차는 2시간 지원됩니다.

3. 발표자: 박예찬 박사(고등과학원)

4. 주요내용: Understanding and Acceleration of Grokking Phenomena in Learning Arithmetic Operations via Kolmogorov-Arnold Representation

We propose novel methodologies aimed at accelerating the grokking phenomenon, which refers to the rapid increment of test accuracy after a long period of overfitting as reported in~\cite{power2022grokking}. Focusing on the grokking phenomenon that arises in learning arithmetic binary operations via the transformer model, we begin with a discussion on data augmentation in the case of commutative binary operations. To further accelerate, we elucidate arithmetic operations through the lens of the Kolmogorov-Arnold (KA) representation theorem, revealing its correspondence to the transformer architecture: embedding, decoder block, and classifier. Observing the shared structure between KA representations associated with binary operations, we suggest various transfer learning mechanisms that expedite grokking. This interpretation is substantiated through a series of rigorous experiments. In addition, our approach is successful in learning two nonstandard arithmetic tasks: composition of operations and a system of equations. Furthermore, we reveal that the model is capable of learning arithmetic operations using a limited number of tokens under embedding transfer, which is supported by a set of experiments as well.

유튜브 스트리밍 예정입니다.