BEGIN:VCALENDAR VERSION:2.0 PRODID:-//https://techplay.jp//JP CALSCALE:GREGORIAN METHOD:PUBLISH X-WR-CALDESC:[24th AIP Open Seminar] Talks by Search and Parallel Computing Unit X-WR-CALNAME:[24th AIP Open Seminar] Talks by Search and Parallel Computing Unit X-WR-TIMEZONE:Asia/Tokyo BEGIN:VTIMEZONE TZID:Asia/Tokyo BEGIN:STANDARD DTSTART:19700101T000000 TZOFFSETFROM:+0900 TZOFFSETTO:+0900 TZNAME:JST END:STANDARD END:VTIMEZONE BEGIN:VEVENT UID:815206@techplay.jp SUMMARY:[24th AIP Open Seminar] Talks by Search and Parallel Computing Unit DTSTART;TZID=Asia/Tokyo:20210512T150000 DTEND;TZID=Asia/Tokyo:20210512T170000 DTSTAMP:20260407T224944Z CREATED:20210419T060011Z DESCRIPTION:イベント詳細はこちら\nhttps://techplay.jp/event/81520 6?utm_medium=referral&utm_source=ics&utm_campaign=ics\n\nSearch and Paral lel Computing Unit (https://aip.riken.jp/labs/generic_tech/search_parall elcomput/?lang=en) at RIKEN AIP\n\nSpeaker 1: (10 min.) Kazuki Yoshizoe \ nTitle: Overview of Search and Parallel Computing Unit\n\nSpeaker 2: (25 min.) Vidal Alcázar\nTitle: A Small Overview about Heuristic Search and Recent Developments in Bidirectional Search.\nAbstract: \nAlready in the first AI conference ever\, at Dartmouth in 1956\, the main paradigm used to solve problems like winning a game or proving a theorem was called "re asoning as search". This consists of advancing towards the goals as if th rough a maze. In order to avoid getting lost in the maze\, computers will use heuristics\, aids that help the computer make informed decisions whe n navigating the search space.\n\nDespite how old and how general heurist ic search is as an approach\, it remains a relatively small and somehow u nknown field. Indeed\, most people wouldn't think of heuristic search whe n discussing milestones in AI such as Deep Blue and AlphaZero\, and yet h euristic search is at the core of both of them.\n\nIn this talk\, we intr oduce the field of heuristic search and its relationship with general int elligence\, and show how even after more than 50 years of important advan cements are being discovered\, like recent breakthroughs in bidirectional search.\n\nSpeaker 3: (25 min.) Kazuki Yoshizoe\nTitle: Massively Paral lel Monte-Carlo Tree Search and Application to Molecular Design\nAbstract : \nIt is common practice to use large computational resources to train n eural networks\, known from many examples\, such as reinforcement learnin g applications. However\, while massively parallel computing is often use d for training models\, it is rarely used to search solutions for combina torial optimization problems.\nOften it is misunderstood that large-scale parallel search is impossible. In this talk\, we describe our past work on 1\,000 worker scale parallel Monte-Carlo Tree Search (MCTS) and propos e an improved massively parallel Monte-Carlo Tree Search (MP-MCTS) algori thm. MP-MCTS works efficiently for a 1\,000 worker scale on a distributed memory environment using multiple compute nodes and we apply it to molec ular design.\n\nThis is the first work that applies distributed MCTS to a real-world and non-game problem. Existing works on large-scale parallel MCTS show efficient scalability in terms of the number of rollouts up to 100 workers. Still\, they suffer from the degradation in the quality of t he solutions. MP-MCTS maintains the search quality at a larger scale. By running MP-MCTS on 256 CPU cores for only 10 minutes\, we obtained candid ate molecules with similar scores to non-parallel MCTS running for 42 hou rs. Moreover\, our results based on parallel MCTS (combined with a simple RNN model) significantly outperform existing state-of-the-art work. Our method is generic and is expected to speed up other applications of MCTS. \n\n(Main part of this talk is presented at ICLR 2021.)\n\nSpeaker 4: (2 5 min.) Kazuki Yoshizoe\nTitle: Massively Parallel Statistical Pattern Mi ning\nAbstract: \nDiscovering significant feature combinations from a lar ge number of features is an important problem in data mining and has vari ous applications in many domains. Naive algorithms fail because the compl exity of the problem is exponential to the number of features. We efficie ntly enumerate statistically significant combinations of features using a massively parallel depth-first search algorithm.\n\nOur statistical patt ern mining uses frequent itemset mining (also known as the "beer diaper p roblem") with pruning based on statistical significance (such as Fisher P -value). We describe how a depth-first search realizes this and how we pa rallelize it at a massive scale. The proposed algorithm\, Massively Paral lel Limitless Arity Multiple-testing Procedure (MP-LAMP)\, works efficien tly on 10\,000 core scale (or larger) clusters.\n\nWe also show two appli cations of MP-LAMP to real-world problems. One is a Genome-Wide Associati on Study (GWAS)\, which analyzes variants of DNAs associated with a trait (such as a genetic disease). Another is a discovery of hidden risk combi nations of posttraumatic stress disorder symptoms.\n\nSpeaker 5: (25 min .) Ryuichiro Hataya\nTitle: Faster AutoAugment: differentiable data augme ntation search for image\nrecognition\nAbstract: \nData augmentation meth ods are indispensable heuristics to improve the performance of deep neura l networks\, especially in image recognition tasks. Recently\, several st udies\, such as AutoAugment [Cubuk et al. 2019]\, have\nshown that augmen tation strategies found by search algorithms outperform hand-made strateg ies. Such methods employ black-box search algorithms over image transform ations with continuous or discrete parameters and require a long time to obtain better strategies.\n\nIn this talk\, we introduce our proposed met hod\, Faster AutoAugment [Hataya et al. 2020]. This method uses a differe ntiable policy search pipeline for data augmentation\, which is much fast er than previous methods. For this purpose\, we used approximate gradient s for several transformation operations with discrete parameters and a di fferentiable mechanism for selecting operations. As the objective of trai ning\, we proposed to minimize the distance between the distributions of augmented and original data\, which can also be differentiated. Image cla ssification experiments show that our method achieves significantly faste r searching than prior methods without a performance drop.\n\n\n\nAll par ticipants are required to agree with the AIP Open Seminar Series Code of Conduct.\nPlease see the URL below.\nhttps://aip.riken.jp/event-list/term sofparticipation/?lang=en\n\nRIKEN AIP will expect adherence to this code throughout the event. We expect cooperation from all participants to hel p ensure a safe environment for everybody.\n\n LOCATION:オンライン URL:https://techplay.jp/event/815206?utm_medium=referral&utm_source=ics&utm _campaign=ics END:VEVENT END:VCALENDAR