Fujitsu World Tour Fujitsu North America Technology Forum Intel Inside. Powerful Productivity Outside. shaping tomorrow with you

Size: px

Start display at page:

Download "Fujitsu World Tour Fujitsu North America Technology Forum Intel Inside. Powerful Productivity Outside. shaping tomorrow with you"

Magdalen Bond
6 years ago
Views:

Fujitsu World Tour 2017 shaping tomorrow

2 Connecting AI Technologies to Real-World Needs Dr. Fumihiro Maruyama Senior Expert Artificial Intelligence Research Center FUJITSU LABORATORIES LTD. Intel Inside. Powerful Productivity Outside. 1 Copyright 2017 FUJITSU

3 Agenda 1. FUJITSU AI Technology Brand Zinrai 2. Cutting-edge AI technologies being developed by Fujitsu Laboratories 3. Connecting AI Technologies to Real-World Needs 2 Copyright 2017 FUJITSU LABORATORIES LTD.

4 Fujitsu s AI goals To create AI that collaborates with people and is human centric To create AI that continuously evolves To create AI that can be incorporated in products and services then deployed 3 Copyright 2017 FUJITSU LABORATORIES LTD.

innovation in business and society by rapidly supporting

5 FUJITSU AI Technology Brand Meaning Being agile and intense Concept behind the brand name Dynamically realizing innovation in business and society by rapidly supporting decision making and actions 4 Copyright 2017 FUJITSU LABORATORIES LTD.

& discovery - Pattern discovery - Inference & & Planning - Prediction & & optimization - Interactivity & & recommendation Learning Advanced

6 A comprehensive framework for AI Sensing People / Businesses / Society Human Centric AI Zinrai Actuation Sensing and Recognition Knowledge Processing Decision and support - Image - Voice - Emotion/state - Natural-language - Natural-language processing - Knowledge processing processing & discovery - Pattern discovery - Inference & & Planning - Prediction & & optimization - Interactivity & & recommendation Learning Advanced research Deep p l e learning a i n g Machine learning Reinforcement learning Neuroscience Social receptivity Simulation 5 Copyright 2017 FUJITSU LABORATORIES LTD.

We extend the scope of Deep Learning Deep

learning that extracts knowledge from data

issue of machine learning Fujitsu Laboratories

Vehicle Recognition Company A: White Practical

Leveraged by topological data analysis Graph

8 We extend the scope of Deep Learning Deep Learning Breakthrough technology of machine learning that extracts knowledge from data Automates feature extraction, a fundamental issue of machine learning Fujitsu Laboratories scope Conventional scope Image Voice Text Vehicle Recognition Company A: White Practical application Person Recognition male Age: 30 Cloth: Gray Backpack : Pink Handwriting Recognition Time-series data Original (2016/02) Leveraged by topological data analysis Graph data Original (2016/10) Based on tensor expression 7 Copyright 2017 FUJITSU LABORATORIES LTD.

o d a t G y r a Betti 数 70 60 50 40 30 20 10 エレベーター内での移動 A エレベーター内での移動 B ランニングマシーンA ランニングマシーンB ステッパーマシーンA ステッパーマシーンB 0 0 100 200 300 400 500 600 Classification of Time-Series Data It is difficult to

Fujitsu has developed deep learning technology that uses advanced chaos theory and topology to automatically and accurately classify erratic time-series data. 1 0.5 0-0.5-1 -1.

9 o d a t G y r a Betti 数エレベーター内での移動 A エレベーター内での移動 B ランニングマシーンA ランニングマシーンB ステッパーマシーンA ステッパーマシーンB Classification of Time-Series Data It is difficult to accurately classify time-series data with extreme oscillations, such as that received from IoT devices. Fujitsu has developed deep learning technology that uses advanced chaos theory and topology to automatically and accurately classify erratic time-series data エレベーター内での移動 A B ランニングマシーン B ステッパーマシーン A B Gyro sensor data Time[s] Classification Accuracy: 85% (25% better than conventional methods) *UC Irvine Machine Learning Repository benchmark test Movement A within elevator Movement B within elevator Treadmill A Treadmill B Stair-stepper A Stair-stepper B Step 1: Graphically represents time-series data using chaos theory (Attractor) Step 2: Quantifies diagrams using topology (Topological Data Analysis) 半径 Step 3: Learning and classification using a convolutional neural network To be incorporated into an IoT product in FY2016. Other time-series data includes log data from servers and sensor data obtained during sleep. 8 Copyright 2017 FUJITSU LABORATORIES LTD.

Classification of Time-Series Data Demo (

Deep Learning with graph-structured data Purpose Technology To elicit new insights from graph-structured data that expresses connections between people and things Technology that allows existing deep

11 Deep Learning with graph-structured data Purpose Technology To elicit new insights from graph-structured data that expresses connections between people and things Technology that allows existing deep learning to be applied to graph-structured data New tensor factorization technology converts graph-structured data to a uniform expression Optimizing uniform expressions and neural network learning simultaneously Tensor-based uniform expression including whole structure グラフ全体構造を含んだ統一的表現 ( テンソル表現 ) に変換 Class A Class B Tensor factorization 統一的表現の最適化 Wider variety of features than handmade (red) 人手の設計 ( 赤 ) より多様な特徴量を自動生成 Extended backpropagation Conventional backpropagation 10 Copyright 2017 FUJITSU LABORATORIES LTD.

12 Graph data appear in various business domains Finance Supply Chain Telecom/Mobile Biology Graph data that expresses relationships between people and things Healthcare/Drug Discovery Social Network Security Internet 11 Copyright 2017 FUJITSU LABORATORIES LTD.

IT drug discovery: virtual screening Demo Predicting a bond between a

14 Outcomes in various domains Outcome Discovered new insights and greatly improved learning accuracy Fintech IoT Medical/Drug Account transaction Credit risk prediction Detect high-risk loans False positive rate: About 20% reduction Benchmark Open Data Institute UK P2P lending market dataset Communication log Intrusion detection Detect intrusions via unexpected paths False positive rate: About 30% reduction Benchmark DARPA Intrusion Detection Evaluation Data Set IT drug discovery Virtual screening Discovered about 200 new protein-compound bond features Activity prediction accuracy: About 10% increase Benchmark PubChem BioAssay data set 13 Copyright 2017 FUJITSU LABORATORIES LTD.

Conversion Linked Open Data (LOD) Large-scale data storing and searching technology for LOD Developed and

search LODs on the web New applications utilizing disparate data Open to the public since January 2014

net Visualizing characteristics of the region Open to the public since December 2014 Convert and

15 Conversion Linked Open Data (LOD) Large-scale data storing and searching technology for LOD Developed and published LOD4ALL service based on fast graph database engine in collaboration with INSIGHT Store and search LODs on the web New applications utilizing disparate data Open to the public since January 2014 Platform for the contest LOD Collaboration with INSIGHT EvaCva Visualizing characteristics of the region Open to the public since December 2014 Convert and integrate data into linked data Store Large-scale Graph DB Attach links RDF Store Text 14 Copyright 2017 FUJITSU LABORATORIES LTD.

Corporate analysis (HighStreet) Purpose Technology Cross-sectional analysis of various aspects of corporate activities from financial statements,

detect similarities and anomalies Outcome Discover discrepancies by integrating multiple reports Visualize relations between enterprises via people by

16 Corporate analysis (HighStreet) Purpose Technology Cross-sectional analysis of various aspects of corporate activities from financial statements, stock prices, products and reputations Integrating various data including financial and social network information Dynamic graph network analysis to detect similarities and anomalies Outcome Discover discrepancies by integrating multiple reports Visualize relations between enterprises via people by integrating executive lists Discovery of new relationships beyond performance comparison and business transactions 15 Copyright 2017 FUJITSU LABORATORIES LTD.

AMED: Integrated Database for Clinical Genome Fujitsu and Kyoto University Start R&D for

knowledge base that aggregates academic paper and public databases in the medical field,

the unique machine learning technology PubMed Variant DB ClinVar Drug DB drugbank

Pathogenic Function Disease Drug LOD Knowledge Base AI Curation@11 hospital

$Learning Clinical Info. Cancer Dementia Genome Info. Refractory disease Infection etc.$

17 AMED: Integrated Database for Clinical Genome Fujitsu and Kyoto University Start R&D for AI that Infers Clinical Interpretations of Genome and Gene Polymorphism Building a knowledge base that aggregates academic paper and public databases in the medical field, using LOD utilization platform Inferring clinical interpretations of Gene Variant using the unique machine learning technology PubMed Variant DB ClinVar Drug DB drugbank Aggregate several databases Academic Paper Original DB LOD Search Engine Gene Protein Pathogenic Function Disease Drug LOD Knowledge Base AI Curation@11 hospital Interpretations Inference Evidence Candidate Treatments Inference Engine based on Machine Learning Clinical Info. Cancer Dementia Genome Info. Refractory disease Infection etc. Integration Integrated Database for Clinical Genome To achieve Genome Medicine 16 Copyright 2017 FUJITSU LABORATORIES LTD.

Collaborative Research with San Carlos Hospital Purpose

clinical decision-making HIKARI Advanced Clinical

combining medical records and public open data

clinical and non-clinical data sources for ensuring

HIKARI Non-clinical data Compare results Outcome 30

years of experience) ~ 20 minutes per patient HIKARI

18 Collaborative Research with San Carlos Hospital Purpose Technology New healthcare solution designed to improve clinical decision-making HIKARI Advanced Clinical Research Information System Knowledge discovery by combining medical records and public open data Anonymization of integrated data gathered from both clinical and non-clinical data sources for ensuring patient privacy Less than 5 seconds 36,000 patients HIKARI Non-clinical data Compare results Outcome 30 patients Clinical history 5 senior clinicians (18-25 years of experience) ~ 20 minutes per patient HIKARI gives clinicians more time and better information to help the patient 17 Copyright 2017 FUJITSU LABORATORIES LTD.

Interactive Q&A Purpose Technology Outcome I want to go to Hawaii.

interaction by utilizing LOD I will go to Hokkaido, where my sister lives, with a friend. Hokkaido. It is still cold there.

answer candidates with relevant information Selecting an answer Go Sister Friend Companion Go Wish Knowledge Selection Rule Going to Hawaii is a wish.

19 Interactive Q&A Purpose Technology Outcome I want to go to Hawaii. Extract necessary information through natural interaction with the user Accurate information extraction by analyzing relations between words Realizing natural interaction by utilizing LOD I will go to Hokkaido, where my sister lives, with a friend. Hokkaido. It is still cold there. Speech Understanding Semantic Structure Analysis Intention Detection Hokkaido Place Lives Actor Hawaii Not destination Knowledge-based Dialog Generation Generating answer candidates with relevant information Selecting an answer Go Sister Friend Companion Go Wish Knowledge Selection Rule Going to Hawaii is a wish. The real destination is Hokkaido. The companion is not a sister but a friend Machine Learning Incorporate Machine Learning Interaction Data LOD Interaction History The customer highly appreciated the merit of the technology based on a trial Looking at deployment in customer-facing business and sales 18 Copyright 2017 FUJITSU LABORATORIES LTD.

AI is relevant to every business AI Corporate Analysis Proxy Dialogue Fraud Detection Finance AI Contamination Test Product Line Automation Symptom Inspection Manufacturing AI Behavior Analysis

22 AI is relevant to every business AI Corporate Analysis Proxy Dialogue Fraud Detection Finance AI Contamination Test Product Line Automation Symptom Inspection Manufacturing AI Behavior Analysis Recommendation Automatic Ordering Retail AI Diagnosis Dialogue Assistance Automated Operation Social Infrastructure Services AI Drug Design State Estimation Image Diagnosis Healthcare AI Fault Diagnosis Advanced Industrial Robot Image-based Decision Monozukuri Congestion Mitigation AI Car AI Self-Driving Video Recognition IoT Business AI Scene Support Drone Wearable Function Installation Device Embedding AI SDN Optimization Automated Operation Fault Recovery Network Device Embedding Function Automated Operation Fault Recovery Installation AI Cloud/ Middleware AI Automated Operation Fault Recovery Server/ Data center 21 Copyright 2017 FUJITSU LABORATORIES LTD.

23 Issues and Approaches Issues Bridging gaps between (general-purpose) AI technologies and real-world needs Streamlining and accelerating of application processes Ongoing approaches PoCs/PoBs with customers Sharing knowledge about customers inside Fujitsu Providing application-oriented APIs in addition to function-oriented ones 22 Copyright 2017 FUJITSU LABORATORIES LTD.

24 Zinrai Platform Service 30 APIs will be available (some are from April 2017) Function-oriented APIs for basic use and application-oriented APIs for rapid development Proven through more than 300 AI-related business projects and field trials Sensing & Recognition Image Speech Emotion processing processing and state Knowledge Processing Naturallanguage processing Knowledge processing and discovery Pattern discovery Decision & Support Inference Prediction & Conversations & recommen and planning optimization -dations Image Speech to text Emotion Text analysis Knowledge structuring Classification Inference Prediction Dialog generation Functionoriented APIs Handwrittentext Video Speech synthesis Line-of-sight Behavior State Speech comprehension Knowledge searching Discrimination Optimization Question- answering Matching Applicationoriented APIs Traffic image Company Domain specific information semantic search searching Credit scoring Demand prediction Delivery planning Production and FAQ search logistics network Communi -cations bot 23 Copyright 2017 FUJITSU LABORATORIES LTD.

25 Fujitsu s Advanced Technologies Applied to APIs Image processing with the world's fastest class of deep learning processing High-precision technology for context comprehension Automatic, natural dialog generation The world's largest scale LOD search Data linking technology with high accuracy Sensing & Recognition Knowledge Processing Decision & Support Image processing Speech processing Emotion and state Naturallanguage processing Knowledge processing and discovery Pattern discovery Inference and planning Prediction & optimization Conversations & recommen -dations Image Speech to text Emotion Text analysis Knowledge structuring Classification Inference Prediction Dialog generation Functionoriented APIs Handwritten- text Video Speech synthesis Line-of-sight Behavior State Speech comprehension Knowledge Knowledge searching searching Discrimination Optimization Matching Question- answering Applicationoriented APIs Traffic image Domain Specific specific Semantic semantic Search search Company information searching Credit scoring Demand prediction Delivery planning Production and logistics network FAQ Search search Communi -cations bot 24 Copyright 2017 FUJITSU LABORATORIES LTD.

Fujitsu Laboratories Advanced Technology Symposium 2018

Fujitsu Laboratories Advanced Technology Symposium 2018 October 9, 2018 Trust and Co-creation in the Digital Era Shigeru Sasaki Fujitsu Laboratories Ltd. CEO Fujitsu Limited CTO 2 FLATS 2017 Quantum Computing: