XAIENCE 2019

7 Nov 2019, 08:00 → 8 Nov 2019, 20:00 ROK

국제회의실 (International Conference Hall) (Seoul National University, College of Natural Sciences, Building 25-1)

The 1st XAIENCE Conference on

'Crossing-over the AI and Science'

XAIENCE 2019 HOME

Speakers

Travel Information

 

 

Slides xaience2019_poster_final.jpg

Thursday, 7 November

08:00 → 09:00 Registration and reception

09:00 → 09:20 Welcome remarks on the 1st XAIENCE conference

Convener: Hyung Do Kim (Seoul National University)

09:00 Welcome remarks on the 1st XAIENCE Conference

Speaker: Prof. Junho Lee (Seoul National University)

09:20 → 10:00 Invited talks - Math & Stats 1

Convener: Prof. Yongdai Kim (Seoul National University)

09:20 Theoretical Advantages of Deep Neural Networks

From statistical points of view, deep neural networks (DNN) are nothing but a (generalized) regression model, but DNNs have solved many problems no other methods have not succeeded in the past. In this talk, I will explain theoretical advantages of DNNs compared to other nonparametric regression models.

Speaker: Prof. Yongdai Kim (Seoul National University)

Slides 1107_MATHSTAT_1_talk_Yongdai_Kim.pdf

09:40 Projection onto Minkowski Sums with Application to Constrained Learning

We introduce block descent algorithms for projecting onto Minkowski sums of sets. Projection onto such sets is a crucial step in many statistical learning problems, and may regularize complexity of solutions to an optimization problem or arise in dual formulations of penalty methods. We show that projecting onto the Minkowski sum admits simple, efficient algorithms when complications such as overlapping constraints pose challenges to existing methods. We prove that our algorithm converges linearly when sets are strongly convex or satisfy an error bound condition, and extend the theory and methods to encompass non-convex sets as well. We demonstrate empirical advantages in runtime and accuracy over competitors in applications to ℓ1,p-regularized learning, constrained lasso, and overlapping group lasso. Reference: 1. Joong-Ho Won, Jason Xu, Kenneth Lange; [Proceedings of the 36th International Conference on Machine Learning, PMLR 97:3642-3651, 2019][1]. [1]: http://proceedings.mlr.press/v97/lange19a.html

Speaker: Prof. Joong-Ho Won (Seoul National University)

Slides https://icml.cc/media/Slides/icml/2019/103(11-14-00)-11-14-00-4664-projection_onto.pdf

10:00 → 10:20 Coffee break

10:20 → 11:50 Invited talks - Earth and Environmental Sciences 1

Convener: Prof. Duk-jin Kim (Seoul National University)

10:20 Lessons Learned from Running Interdisciplinary Computational Sciences Program for Undergraduate at Seoul National University on the Future of Scientific Computing Education

자연과학은 자연을 대상으로 한 관측이 주된 목표이고 우리 자연과학자들은 관측자료 즉 데이터를 통해 자연(모델)을 이해하는 것에 매우 익숙하다. 그런데 최근 갑자기 데이터과학, 인공지능이며 머신러닝이 4차 산업혁명이라는 buzzword와 함께 큰 화제이다. 기업(enterprise) 차원에서는 이것들이 새로운 분야이자 기회인 것 처럼 보일지 모르지만 아직 자연과학 분야에서의 기여도는 제한적이다. 정말 인공지능 머신러닝을 통해 새로운 자연현상이 밝혀진 사례는 최소한 내가 속한 지질해양분야에서는 그리 많은 것 같지 않다. 그럼에도 불구하고 학부생들과 대학원생들에게 새로운 분야를 가르쳐야하는 교육자로써 지난 5-6 년간 교육현장에서 느낀 점들을 공유하고자 한다. 특히 enterprise AI와 달리 scientific AI가 나아가야 할 방향 그리고 컴퓨터과학의 발달로 인해 달라진 ICT의 생태계 속에서 어떻게 미래문제해결형 학생을 양성하기 위한 교육이 이루어져야 하는가에 대해 의견을 제시하고자 한다. 이는 현재 서울대 계산과학 연합전공과 협동과정이 추구하는 방향이기도 하다. Science is all about observation and measurement. So for natural scientists like us we should be very comfortable about how to derive our understanding into models from observed data sets. This should be very familiar to us and It is the crux of what we do day to day. Recently, however, we are puzzled by the new buzzword ‘The Fourth Industrial Revolution’ as well as terms such as data science, machine learning and artificial intelligence. However, at least in my field of marine and Earth sciences, the impact brought about by these new disciplines and approach have been somewhat limited because in science we never get enough data as we want. These new approaches require immense amounts of data, which sometimes is not practical. As a person who has been involved as an educator in the midst of these new changes, I would like to share my views on how we can the important differences between enterprise AI and scientific approaches. It is important to teach students probably under these new changes because our goal has always been to raise competent next-generation scientists to deal with future problems. I hope I can suggest a better picture of scientific computation in this world of AI which we had computational sciences program of Seoul National University is also trying to find.

Speaker: Prof. Sang-Mook Lee (Seoul National University)

Slides 1107_EE_1_talk_Sang-Mook_Lee.pdf

10:50 Deep Learning for Sea Ice Classification and Ship Detection using Satellite Synthetic Aperture Radar

Speaker: Prof. Duk-jin Kim (Seoul National University)

Slides 1107_EE2_XAIENCE_Duk-jin_Kim.pdf

11:10 Deep learning for multi-year ENSO forecasts

Variations in the El Niño/Southern Oscillation (ENSO) are associated with a wide array of regional climate extremes and ecosystem impacts. Robust, long-lead forecasts would therefore be valuable for managing policy responses. But despite decades of effort, forecasting ENSO events at lead times of more than one year remains problematic. Here we show that a statistical forecast model employing a deep-learning approach produces skilful ENSO forecasts for lead times of up to one and a half years. To circumvent the limited amount of observation data, we use transfer learning to train a convolutional neural network (CNN) first on historical simulations and subsequently on reanalysis from 1871 to 1973. During the validation period from 1984 to 2017, the all-season correlation skill of the Nino3.4 index of the CNN model is much higher than those of current state-of-the-art dynamical forecast systems. The CNN model is also better at predicting the detailed zonal distribution of sea surface temperatures, overcoming a weakness of dynamical forecast models. A heat map analysis indicates that the CNN model predicts ENSO events using physically reasonable precursors. The CNN model is thus a powerful tool for both the prediction of ENSO events and for the analysis of their associated complex mechanisms. References: [https://www.nature.com/articles/s41586-019-1559-7][1] [https://www.sciencemag.org/news/2019/09/artificial-intelligence-could-predict-el-ni-o-18-months-advance][2] [https://www.yna.co.kr/view/AKR20190918093900017?input=1195m][3] [http://biz.chosun.com/site/data/html_dir/2019/09/19/2019091900659.html][4] [1]: https://www.nature.com/articles/s41586-019-1559-7 [2]: https://www.sciencemag.org/news/2019/09/artificial-intelligence-could-predict-el-ni-o-18-months-advance [3]: https://www.yna.co.kr/view/AKR20190918093900017?input=1195m [4]: http://biz.chosun.com/site/data/html_dir/2019/09/19/2019091900659.html

Speaker: Prof. Yoo-Geun Ham (Chonnam National University)

Slides 1107_EE3_XAIENCE_Yoo-Geun_Ham.pdf

11:30 AI Applications for Satellite-based Disaster Monitoring and Prediction

Speaker: Prof. Jungho Im (UNIST)

Slides 1107_EE_4_Jungho_Im_dist.pdf

11:50 → 13:30 Lunch

13:30 → 14:30 Invited talks - Chemistry 1

Convener: Prof. Chaok Seok (Seoul National University)

13:30 Prediction of Protein Structure and Interaction by Physics and Informatics

Protein structure prediction problem has challenged theoretical and computational physical scientists since the first protein structure was published in 1958. There have been steady progresses in protein structure prediction since then, but major contributions to the progress came from informatics-based approaches rather than from physics-based approaches. Recently, DeepMind’s AlphaFold made a further contribution by introducing deep learning to extract structural information from the large sequence database. Meaningful contribution of physics-based approaches began to be made only in 2012 in the field of structure refinement. However, structural improvements that can be achieved by refinement with current physics-based approaches are very limited due to both energy and sampling problems. To overcome this limitation, we are taking an approach that combines physics and informatics, including deep learning. We take similar approaches to predict interactions of proteins with other proteins or small ligands including short peptides and oligosaccharides. Our goal is to develop protein structure modeling techniques that can provide useful predictions even in the absence of available information, although currently available experimental data would play important roles in developing such techniques. Such modeling methods would be very useful for applications to a wide range of biomedical research and drug discovery.

Speaker: Prof. Chaok Seok (Seoul National University)

Slides 1107_CHEM_1_talk_Chaok_Seok.pdf

13:50 Discovering Novel Fluorescent Molecules by Combining Machine-learning and Global Optimization

Fluorescent molecules are widely used for bio-imaging. They are attached to specific cell organelles and/or proteins, enabling observation of detailed structure and dynamics in the cell. Efficient fluorescent molecules must have a high quantum yield for effective bio-imaging. Here, we present a systematic approach to discovering novel fluorescent molecules that combines machine-learning and global optimization algorithms. We recast the problem of discovering novel fluorescent molecules with high-intensity emission light into a global optimization problem by using the oscillator strength of a molecule as an objective function for optimization. A statistical machine that predicts excitation energies and associated oscillator strengths, the probability of absorption or emission of light in transitions between different energy states, of a molecule were trained using the random forest algorithm. The Pub-chemQC database, which contains TD-DFT calculation results of 3.8 million known molecules, was used as a training set. The extended connectivity fingerprints of molecules were used as input vectors. To optimize the oscillator strength of a molecule, a highly efficient global optimization algorithm called CSA was used. For CSA global optimization, SMILES representation of a molecule was mapped to a 200-dimensional integer vector by using Natural Language Toolkit. After CSA global optimization calculation converged, we assessed the validity of our approach by performing quantum mechanical calculations. TD-DFT calculations were carried out to verify whether novel molecules obtained by this procedure actually have high oscillator strength.

Speaker: Prof. Juyong Lee (Kangwon National University)

Slides 1107_CHEM_2_talk_Juyong_Lee.pdf

14:10 Fleming : AI-driven Integrated Drug Discovery Platform

New drug development cost more than one billion over the last decade. And the chance of success was quite low, one in five thousands. In order to make the drug development process efficient in terms of time and cost, Artificial Intelligence has been actively introduced to and rigorously adopted in the various fields of drug development: compound activity prediction, compound design, patient selection, and clinical trial design, just to name a few. As these fields become advanced, more AI-based drug development platforms are developed and deployed in the fields to overcome data shortage. Fleming is an integrated and automated platform for time and cost efficient drug development. It uses protein structure prediction and genome-based target selection for efficiency. In Fleming, precise target protein structures, generated by protein structure prediction techniques using AI, are used to pick candidate compounds and to design active compounds with high chance of success. And genome-based analysis powered by Fleming AI selects (find-an-adjective-for-this) compounds by predicting compound activity and toxicity. These features in Fleming accelerate the new drug development process utilizing and combining various technologies: disease genome analysis, candidate compound selection, compound generation, activity prediction, and toxicity prediction.

Speaker: Dr Junsu Ko (© Arontier)

Slides 1107_CHEM_3_talk_Junsu_Ko.pdf

14:30 → 14:50 Coffee break

14:50 → 15:30 Invited talks - Physics and Astronomy 1-1

Convener: Dr Min-Su Shin (KASI)

14:50 Application of Machine Learning for Big Data Analysis in Astronomy (천문학 자료분석에서의 기계학습의 활용)

천문학 관측 자료의 분석의 경우, 관측 환경의 변화 등에 의해서 야기된 관측 부정확성이나 편향성, 그리고 비균질적인 자료 획득 등의 특징적인 문제를 가지고 있다. 이러한 천문 관측 자료의 특성을 고려해서, 천문학자들은 천체나 천체 현상의 검출, 분류, 특성의 추론 등의 문제를 접하게 된다. 이러한 문제들은 관측 자료의 급속한 증가와 함께 더욱 심각해지는데, 기계학습의 활용을 통하여 정량적인 검출, 분류, 추론의 문제들을 해결해 가고자 노력 중이다. 이번 발표에서는 이러한 문제들을 clustering, anomaly detection, classification, regression, recommendation 문제의 시각에서 기계학습을 활용한 사례들을 소개하고, 최근 시도되고 있는 ensemble learning, multi-task learning 등의 활용도 간랸히 제시하고자 한다. 더불어서 천체와 그 현상에 대한 천체물리적 이해를 위한 기계학습의 활용 방향에 대해서 의견을 제시한다.

Speaker: Dr Min-Su Shin (KASI)

Slides 1107_AST_1_talk_Min-Su_Shin.pdf

15:10 Machine Learning In Astrophysics: Estimating Galactic Baryonic Properties from Their Dark Matter

As astrophysics and cosmology deal with inherently "cosmological" sizes of data, machine learning is being rapidly adopted in a variety of astrophysical applications. In this talk, I will introduce a pipeline that estimates baryonic (visible) properties of a galaxy based purely on dark matter (DM; invisible) properties in large-scale DM-only simulations. It is shown that our pipeline promptly generates a galaxy catalogue from a DM halo catalogue using a machine trained on smaller-scale, fully-hydrodynamic, high-resolution simulations. An extremely randomized tree algorithm is used together with multiple novel improvements we developed such as a refined error function and two-stage learning. Our model may become a promising method to transplant the baryon physics of galaxy-scale hydrodynamic calculations onto a larger-volume DM-only run.

Speaker: Prof. Ji-hoon Kim (SNU)

Slides XAIENCE_SNU_110719.pdf

15:30 → 15:45 Coffee break

15:45 → 17:15 Invited talks - Physics and Astronomy 1-2

Convener: Prof. Hyung Do Kim (Seoul National University)

15:45 What Can We Learn with ML in Particle Physics, and How?

ML has a deep potential to extend our ability to understand Nature, beyond common knowledge. As the very first steps to realize it, we use ML to seek for answers to one of the not-well-solved problems in particle physics. Our focus is not only to improve the solution, but to develop ways to figure out what the network has learned.

Speaker: Prof. Sunghoon Jung (Seoul National University)

16:05 Searching for Axions and new concepts in fundamental physics with Machine Learning

In this talk, I will review how ultralight axion-like particles can be constrained using X-ray observations of bright localized sources (AGNs, Quasars) in and behind galaxy clusters. To find axion-like particles in these settings corresponds to finding their characteristic pattern (quasi-sinusoidal oscillations) in noisy data. Using ML-techniques we are able to improve the search sensitivity for these particles and are able to set stronger bounds compared to previous methods. In the second part of the talk, I shall highlight avenues where ML can be used to accelerate our understanding of fundamental physics (string theory) and potentially vice versa.

Speaker: Dr Sven Krippendorf (Ludwig Maximilian University of Munich)

16:35 Learning QCD Jet Flavors in search for Invisible Higgs Decays

Speaker: Prof. Hyung Do Kim (Seoul National University)

Slides 2019XAI_hdkim.pdf

16:55 Recent Developments of Deep Learning in Experimental High Energy Physics

최근들어 딥러닝 기술은 기초 과학 분야에도 광범위하게 연구되고 있다. 이를 통해 얻어진 지식 및 기술이 차세대 연구 시스템 및 환경을 변화시키고 결과를 획기적으로 향상시키기 위해 다양하게 적용될 것으로 기대되고 있다. 현재 고에너지 물리 실험분야에서 활발하게 연구되고 있는 딥러닝 관련 연구에 대해 소개하고 향후 기대 효과에 대해 논의 한다.

Speaker: Prof. Hwidong Yoo (Seoul National University)

Slides HwidongYoo_XAIENCE_11072019.pdf

17:15 → 17:45 Contributed talks 1

Convener: Prof. Yongdai Kim (Seoul National University)

17:15 On Casting Importance Weighted Autoencoder to an EM Algorithm to Learn Deep Generative Models

We propose a new and general approach to learn deep generative models. Our approach is based on a new observation that the importance weighted autoencoders (IWAE) can be understood as a procedure of estimating the MLE with an EM algorithm. Utilizing this interpretation, we develop a new learning algorithm called importance weighted EM algorithm (IWEM). IWEM is an EM algorithm with importance sampling (IS) where the proposal distribution is carefully selected to reduce the variance due to IS. In addition, we devise an annealing strategy to stabilize the learning algorithm. For missing data problems, we propose a modified IWEM algorithm called miss-IWEM. Using multiple benchmark datasets, we demonstrate empirically that our proposed methods outperform IWAE with significant margins for both fully-observed and missing data cases.

Speaker: Dr Dongha Kim (Seoul National University)

Slides 1107_CONT_1_talk_Dongha_Kim.pdf

17:30 Nonconvex Sparse Regularization For Deep Neural Networks and its Optimal Property

Sparsity is a key ingredient in the success of learners both theoretically and computationally. This is also the case for deep neural networks (DNNs). A number of empirical observations show that sparse DNNs can dramatically reduce computation time and memory without appreciably harming prediction power. Furthermore, recent theoretical studies proved that DNN estimators with a certain sparsity constraint can attain optimal convergence rates for regression and classification problems. However, they only considered the empirical risk minimizer under the sparsity constraint, where optimization is almost impossible in practice due to its discrete nature. In this research, we propose a novel penalized empirical risk minimization method for estimating sparse DNNs with a scalable computation algorithm. The proposed method yields sparse DNNs that can achieve optimal convergence rates of excess risks for various learning problems including regression and binary classification. We demonstrate the empirical performance of the proposed method and compare it with other competitors for various benchmark datasets.

Speaker: Mr Ilsang Ohn (Department of Statistics, Seoul National University)

Slides 1107_CONT_2_talk_Ilsang_Ohn.pdf

17:45 → 19:00 Banquet

국제회의실 (강연장 앞 로비)

Friday, 8 November

08:00 → 09:00 Registration and reception

09:00 → 09:40 Invited talks - Biological sciences, Neurosciences

Convener: Prof. Junho Lee (Seoul National University)

09:00 Connectome Project & Neuroscience with AI

예쁜꼬마선충 C. elegans는 신경계의 발생과 진화를 연구할 수 있는 최적의 모델동물이다. 꼬마선충의 신경계는 302개의 뉴런으로 구성된 최소화된 뇌의 모델이라 할 수 있다. 본 연구실에서는 nictation 이라고 하는 히치하이킹 행동의 신경회로의 작동 및 조절, 그리고 진화에 대한 연구를 수행하고 있다. Nictation 행동은 특정 발생 단계 (dauer라는 휴면 단계)의 선충에서만 나타나는 행동으로서 고개를 들어 몸을 세우고 필요한 경우 흔드는 행동을 의미한다. 본 연구실의 선행연구로 많은 뉴런 중 IL2라는 뉴런이 핵심적인 역할을 하는 것을 밝혔고 지구상에 흩어져 있는 다른 꼬마선충들이 그 nictation의 정도를 달리하는 유전적 다양성을 가짐도 보인 바 있다. 현재 가장 중요한 문제는 nictation 행동이 어떻게 dauer라고 하는 휴면 유충에서만 실행되는가이고, 그 답이 될 수 있는 것은 신경회로 즉 커넥톰의 차이에 있을 수 있다는 것이다. 이 가설을 검증하고 신경계의 발생유연성의 기전을 연구하기 위해 우리는 꼬마선충 dauer를 50 나노미터 두께로 연속절편하여 전자현미경 사진을 찍은 후 모든 3차 구조를 재구성하는 작업을 진행하고 있다. 여기서 AI가 관여하는 작업이 가능하다면 이룰 수 있는 일들이 획기적으로 많아질 것으로 기대하며 본 연구의 진척과 문제점 등을 공유하고자 한다.

Speaker: Prof. Junho Lee (Seoul National University)

09:20 Hoping for Better AI: Advice from the Brain

Artificial intelligence (AI), by definition, is an artificial version of natural intelligence (NI). Trying to understand how natural intelligence works, Alan Turing, Demis Hassabis and other AI experts have tried to understand how the human brain works. With the recent explosive advance in technology, modern neuroscience is now getting closer to unraveling the mysterious algorithms of the brain one by one. Some of the algorithmic principles as to how the brain processes information (e.g., reinforcement learning, deep neural networks with backpropagation, etc.) have been readily adopted by the AI research field and have revolutionized the field as evidenced by the famous ‘AlphaGo’ by the Google’s DeepMind. However, the current version of AI is inferior to the brain in many aspects although neuroscience has a lot more to offer in pushing the envelope in AI technology to make AI virtually inseparable from NI in the future. In this short talk, I will introduce some of the key areas in which modern AI to improve to resemble the human brain.

Speaker: Prof. Inah Lee (Seoul National University)

Slides 1108_B2_XAIENCE_Inah_Lee.pdf

09:40 → 10:00 Coffee break

10:00 → 11:00 Invited talks - Math & Stats 2

Convener: Prof. Cheon Jung Hee (Seoul National University)

10:00 데이터의 프라이버시를 보존하는 기계학습 (Safe Machine Learning toward Private AI)

기계학습(Machine Learning)은 데이터로부터 이를 도출한 함수를 유추하는 과정으로 최근 다양한 분야에서 흥미로운 응용들이 제시되고 있다. 기계학습이 좋은 성과를 거두려면 데이터의 확보가 필수적인데 개인 프라이버시 문제 혹은 데이터 주권의 문제로 인해 좋은 데이터를 확보하는 것이 쉽지 않은 일이다. 동형암호는 암호화한 데이터상에서 복호화없이 기계학습의 훈련단계(Training)나 예측단계(Inference)를 수행할 수 있도록 하며, 이를 통해 데이터의 프라이버시 문제를 극복하고 Private AI의 시대를 열어가고 있다. 본 강연에서는 이 분야의 최근 결과들로 동형 회귀분석(Homomorphic Logistric Regression), 동형 심층신경망(Homomorphic Deep Neural Network), 동형 의사결정나무(Homomorphic Decision Tree)등의 결과와 이의 신용정보, 의료, 마케팅 등에의 응용을 소개하도록 한다.

Speaker: Prof. Jung Hee Cheon (Seoul National University)

Slides 1108_M1_XAIENCE_Jung_Hee_Cheon.pdf

10:20 Quantum AI, Quantum Machine Learning

Recent development in quantum technology together with advances in quantum algorithms impacts the field of quantum AI and machine learning. Many quantum machine learning algorithms and their applications to AI are being appeared. We present some of these and also our work in the field.

Speaker: Prof. Dongpyo Chi (Seoul National University)

Slides 1108_M2_XAIENCE_Dongpyo_Chi.pdf

10:40 Harmonic Data Analysis for Shape and Centrality

Harmonic data analysis aims to provide topological and combinatorial summaries of data sets by representing them as simplicial complexes. Topological data analysis, which we shall review briefly, initiated a topological approach and introduced shape of data as a new data scientific feature. Recently, the need for simplicial complexes for data analysis arose again due to the emergence of simplicial networks for modeling higher order relations among data points, which requires both topological insight and combinatorial precision. In this talk, we will present methods from topological combinatorics for refining data shape via harmonic cycles and computing network centrality via simplicial effective conductance. Applications to medicine and social networks will be presented. We will also describe recent experiments in machine learning incorporating mathematical data summary.

Speaker: Prof. Woong Kook (Seoul National University)

Slides 1108_M3_XAIENCE_Woon_Kook.pdf

11:00 → 11:40 Invited talks - Earth and Environmental Sciences 2

Convener: Prof. Duk-jin Kim (Seoul National University)

11:00 Application of Artificial Intelligence techniques to Weather Forecasting

Speaker: Prof. Jung-Hoon Kim (Seoul National University)

Slides 1108_EE1_XAIENCE_Jung_Hoon_Kim.pdf

11:20 Clustering weather maps for high PM10 events in Seoul using Self Organization Maps

Speaker: Prof. Seok-Woo Son (Seoul National University)

Slides 1108_EE2_XAIENCE_Seok_Woo_Son.pdf

11:40 → 13:00 Lunch

13:00 → 13:40 Invited talks - Chemistry 2

Convener: Prof. Chaok Seok (Seoul National University)

13:00 Delfos: deep learning model for prediction of solvation free energies in generic organic solvents

Prediction of aqueous solubilities or hydration free energies is an extensively studied area in machine learning applications in chemistry since water is the sole solvent in the living system. However, for non-aqueous solutions, few machine learning studies have been undertaken so far despite the fact that the solvation mechanism plays an important role in various chemical reactions. Here, we introduce Delfos (deep learning model for solvation free energies in generic organic solvents), which is a novel, machine-learning-based QSPR method which predicts solvation free energies for various organic solute and solvent systems. A novelty of Delfos involves two separate solvent and solute encoder networks that can quantify structural features of given compounds via word embedding and recurrent layers, augmented with the attention mechanism which extracts important substructures from outputs of recurrent neural networks. As a result, the predictor network calculates the solvation free energy of a given solvent–solute pair using features from encoders. With the results obtained from extensive calculations using 2495 solute–solvent pairs, we demonstrate that Delfos not only has great potential in showing accuracy comparable to that of the state-of-the-art computational chemistry methods, but also offers information about which substructures play a dominant role in the solvation process. References: 1. [Delfos: deep learning model for prediction of solvation free energies in generic organic solvents, Hyuntae Lim and YounJoon Jung, Chemical Science 2019 DOI: 10.1039/C9SC02452B (2019)] [1] 2. [Media] [2] [1]: https://pubs.rsc.org/en/content/articlelanding/2019/sc/c9sc02452b#!divAbstract [2]: http://now.snu.ac.kr/47/3/1450

Speaker: Prof. YounJoun Jung (Seoul National University)

Slides 1108_CHEM1_XAIENCE_YounJoon_Jung.pdf

13:20 AI-based Smart Molecular Design

The ultimate goal of chemistry is to make new molecules with desired properties. It is challenging because chemical space is very large and discrete with a wide variety of molecules. For example, there are only 108 molecules synthesized as potential drug candidates, but 1060 molecules are estimated to be existing. High-throughput virtual screening approach has attracted great attention but still requires large costs and time. In this talk, we propose to use a molecular generative model based on deep learning algorithm as an alternative. It is specialized in controlling multiple molecular properties simultaneously, embedding them in namely the latent space. As a proof of concept, we will show that it can be used to generate a number of molecules as drugs with specific properties. We also apply it to design of new molecules with promising binding energy for a specific target protein and use them as potential drug candidates that are not in the database.

Speaker: Prof. Woo Youn Kim (Chemistry, KAIST)

Slides 1108_CHEM2_XAIENCE_Woo_Youn_Kim.pdf

13:40 → 14:00 Coffee break

14:00 → 15:20 Invited talks - Physics and Astronomy 2

Convener: Dr Won Sang Cho (Seoul National University)

14:00 Architecture of Neuromorphic Hardware with Embedded Learning

Speaker: Prof. Doo Seok Jeong (Hanynag University)

Slides 1108_P1_XAIENCE_Doo_Seok_Jeong.pdf

14:20 Information in Unsupervised Learning

Information bottleneck theory explains that neural networks maximally compress unnecessary information in input data, and transfer sufficient information to output. Unlike the supervised learning, the information flow of unsupervised learning has not been explored much. In this talk, I introduce the information extraction of unsupervised learning by using a representative generative model, deep belief networks. Then, I show that the unsupervised learning maximally extracts relevant information given a fixed information compression for its internal representations. This relation between the information bottleneck theory and our theory reminds the duality between the rate distortion theory and channel capacity in information theory. [국문] Information bottleneck 이론은 지도학습과정에서 일어나는 신경망의 정보흐름을 정량화하는 이론이다. 이 이론에 따르면 신경망의 학습은 입력과 출력을 연결시켜주는 최소한의 정보만 전달하고 불필요한 정보는 최대한 압축하는 과정임을 보여주었다. 우리는 입력과 출력이 정의되는 지도학습과 달리 비지도학습과정에서 일어나는 신경망의 정보흐름에 대해 연구했다. 이 발표에서는 대표적인 생성모형인 딥빌리프네트워크를 통해서 비지도학습과정의 정보추출에 대해 소개하려고 한다. 우리는 딥빌리프네트워크가 주어진 해상도의 내적표현을 가질 때, 내적표현의 빈도수에서 최대한의 정보를 추출한다는 사실을 발견하였다. Information bottleneck 이론과 우리 이론의 관계는 마치 정보이론의 rate distortion theory와 channel capacity 사이의 쌍대성을 떠올리게 한다.

Speaker: Prof. Junghyo Jo (Seoul National University)

Slides 1108_P2_XAIENCE_Junghyo_Jo.pdf

14:40 Exploration of energy surfaces and conformation of molecules and solids by utilizing a machine-learning technique

Predicting the physical properties of novel materials requires an accurate description of atomic interactions as provided by first-principles quantum mechanical calculations. Efficient and practical calculation tools have been developed along with the progress of density functional theory (DFT). Still, however, the computational complexity associated with the quantum mechanical treatment limits their applications to systems of a few hundreds of atoms at most. Here, we present an application of the Gaussian process regression (GPR) scheme to the global optimization, conformation space annealing, and pathway optimization methods. We demonstrate that the use of GPR-based pathway optimization technique, e.g., action-derived molecular dynamics (ADMD) method, can be useful in enhancing the computational performance. We will discuss possible future applications of the GPR-based machine learning technique for the exploration of energy surfaces and conformation of molecules and solids.

Speaker: Prof. Jaejun Yu (Seoul National University)

Slides 1108_P3_XAIENCE_Jaejun_Yu.pdf

15:00 Quantum Algorithms and Quantum Machine Learning

정보처리/컴퓨팅에의 초소형화/고집적화 그리고 무어의 법칙 등은 정보이론과 양자물리학의 융합 및 발전이 필연임을 반영한다. 양자물리학에 기반한 컴퓨팅에의 속도향상 등의 연구는 그 가능성을 증명하는 수준을 넘어, 양자기술의 실현이라는 현실적 목표앞에 와 있다. 이제, 양자컴퓨팅/통신 등의 용어는 학계를 넘어 이미 우리 일상에 와 있다. 이에, 본 강연에서는 기본적인 양자원리, 즉 양자중첩 및 얽힘 등을 이해하고, 알려진 유명한 양자알고리즘들과 더불어 어떻게 양자머신러닝에의 속도향상이 가능한지 간략히 살펴보고자 한다. 아울러, 최근 연구 동향 및 이슈 등 또한 간략히 소개하고자 한다.

Speaker: Dr Jeongho Bang (KIAS)

Slides 1108_P4_XAIENCE_Jeongho_Bang.pdf

15:20 → 15:40 Coffee break

15:40 → 16:55 Contributed talks 2

Convener: Dr Won Sang Cho (Seoul National University)

15:40 Probing Trilinear Higgs Self–coupling at the HL-LHC with Machine Learning

1964년, 피터 힉스에 의해 제안된 힉스 메커니즘을 통해 힉스입자의 존재가 예측되었다. 이후 48년이 지나, 2012년 7월 유럽입자물리연구소(CERN) 대형강입자 충돌기(LHC)에서 드디어 이 입자가 발견되었다. 현재 이 입자의 성질을 정밀하게 측정하고 검증하고자, LHC에서 다양한 연구들이 진행 중에 있다. 이 발표에서는 이 중, 힉스입자의 삼중 커플링의 검측 성능을 머신러닝의 기술을 이용해 개선한 예를 통해, 고에너지 물리학에서의 응용을 소개하고자 한다.

Speaker: Prof. Jubin Park (Institute for Universe and Elementary Particles, Chonnam National University)

Slides 1108_CONT1_XAIENCE_Jubin_Park.pdf

15:55 Prediction and Taking Insight of Molecular Quantum Property by Random Forest Regression

Fluorescent molecules are widely used for bio-imaging. They are attached to specific cell organelles or proteins, enabling observation of detailed structure and dynamics in the cell. Efficient fluorescent molecules must have a high quantum yield for effective bio-imaging. Diverse effective fluorescent molecules whose color are distinctive need to get more information of cell or protein. First step of discover novel molecules using computational approach is prediction of compound. Here, we use random forest regression to predict excitation energy and oscillator strength of a molecule. Random forest algorithm is white box. It is easy to extract feature importance. We could get insight from it. A statistical machine that predicts excitation energies and associated oscillator strengths, the probability of absorption or emission of light in transitions between different energy states, of a molecule were trained using the random forest algorithm. The PubchemQC database was used as a training set. It has over 3 million known compounds. We picked up 0.5 million molecules randomly. 90% of them were in training set and the others were test set. The ECFP4(extended connectivity fingerprints 2) of molecules were used as input features. We found some fragments which can decide molecules’ quantum property by feature importance analysis.

Speaker: Mr Beomchang Kang (Seoul National University)

Slides 1108_CONT2_XAIENCE_Beomchang_Kang.pdf

16:10 Predicting location of suitable groundwater for brewing coffee using tree-based ensemble machine learning

Recently, groundwater is used as a source of drinking purposes such as water, coffee, beer and other beverages. The range of water quality for producing high-quality of beverages is different from each usage, so it is important to find suitable groundwater location for each purpose in the aspect of water industry. This study was conducted to predict the suitable location for brewing coffee in Gangwon Province, South Korea using tree-based ensemble machine learning. Appropriate water quality standard for brewing coffee is known as TDS of 75~250 mg/L and calcium hardness of 17~85 mg/L from recent research. Boosted Regression Trees (BRT), Random Forests (RF) and Extremely Randomized Trees (ERT) were used as tree-based ensemble method. Response indicating suitable or unsuitable groundwater for brewing coffee was determined by 254 wells’ water quality data, and predictor variables were composed of slope, altitude, drainage grade, effective soil depth, soil composition, land use, and hydrogeology based on GIS data. Applying models to the test data, all three models showed the area under a curve (AUC) and accuracy more than 0.85 and 0.80, respectively, which indicates high reliability in prediction. Threshold of dividing suitable or unsuitable for brewing coffee is found from Receiver Operating Characteristic (ROC) curve. The BRT showed the highest AUC and accuracy among the three models, therefore, potential map of suitable groundwater location for brewing coffee was suggested by the BRT model. In the condition of lack of water quality data, this research can help to determine location of suitable groundwater for several usages. Keyword: potable groundwater · boosted regression tree · random forest · extremely randomized tree Acknowledgement: This research was supported by the National Research Council of Science and Technology(NST) grant funded by the Korea government(MSIP) (No. CAP-17-05-KIGAM)

Speaker: Ms Hye-Lim Lee (School of Earth and Environmental Sciences, Seoul National University)

Slides 1108_CONT3_XAIENCE_Hye-Lim_Lee.pdf

16:25 Application of Neural Network Model to Predict and to Evaluate the Groundwater Level Fluctuation

Groundwater is one of valuable water resources used for the various purposes in our life, but it is not limitless. For this reason, prediction and management of the groundwater are very essential work for sustainable use of it. Physics-based models are usually applied on prediction of groundwater level, but they are hard to be implemented successfully when there is any unknown physical property or when subterranean structure is very complicated. In this research, prediction of the groundwater level at riverside area in Yangpyeong, Korea was carried out using neural network model instead. In study area where several natural and anthropogenic factors affect the groundwater level fluctuation, groundwater levels at 8 monitoring wells were well predicted with low range of RMSE errors. Moreover, monthly contributions, which indicate the impact of input variables, were computed to figure out the seasonal variance of influencing factors. This study could suggest another option to predict the groundwater level, and help understanding spatial and temporal variation of impacts of factors affecting the groundwater level fluctuation. Acknowledgement: This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIP) (No. 2017R1A2B3002119)

Speaker: Mr Sanghoon Lee (Seoul National University)

Slides 1108_CONT4_XAIENCE_Sanghoon_Lee.pdf

16:40 복소 산술 신경망을 통한 물리적 기계학습모형의 건설과 물리적 불변량의 추출

Speaker: Dr Won Sang Cho (Seoul National University)

Slides 1108_CONT5_XAIENCE_Won_Sang_Cho.pdf

16:55 → 17:20 Concluding remarks

16:55 Adjourn