Artificial Intelligence is here - Amazon S3€¦ · Rapid technology advancements • Unprecedented...
Transcript of Artificial Intelligence is here - Amazon S3€¦ · Rapid technology advancements • Unprecedented...
www.techarchday.fiwww.techarchday.fi
Artificial Intelligence
is here
Jarkko Ylipaavalniemi
D.Sc. (Tech), IEEE CBP
Technology Architect, Accenture
www.techarchday.fi
What is artificial intelligence?
1: a branch of computer science dealing with the simulation of intelligent behavior in computers
2: the capability of a machine to imitate intelligent human behavior
the theory and development of computer systems able to perform tasks normally requiring human intelligence, such as visual perception, speech recognition, decision-making, and translation between languages
www.techarchday.fi
Human intelligence
• The human brain connectome
– It contains around 100 billion neurons
– Each neuron can have up to 10 000 connections to other neurons
– Connections are mediated through an estimated 1 000 trillion synapses
– Perhaps a memory of 1 000 terabytes
– And tons of other stuff too…
The human brain has a peakpower consumption of 20 W
www.techarchday.fi
AI in practice – a typical day?
7 am Gesture Alarm Clock: Alarm clock wakes you up 15 minutes early because it detects traffic delays
8 am Tesla Autopilot: Model S automates steer, turn, speed, park, and data is transmitted to the fleet to help system learn and improve
10 am Charles Schwab’s Robo-Advisor: Allocates stocks and investments based on a completed questionnaire
2 pm Enlitic: Turns large amounts of medical diagnostic data into deep insights and finds subtle patterns
4 pm Cubic: Suggests a movie when you ask “What should I do tonight?”
5 pm Automated Insights: WordSmith finishes writing a news article to be published on Associated Press news in just two minutes
8 pm X.ai: Ask virtual assistant, Amy, to schedule a meeting with CEO at 2pm tomorrow
2 am Keeker: While you are sleeping, smart robot keeps your home secure
www.techarchday.fi
Where are we today?
Apple’s Siri serves over a billion requests a week with 5% word error rate
Facebook’s DeepFace uses a 9-layer deep neural network having over 120 million parameters trained with 4 million facial images belonging to 4.000 identities, reaching an accuracy of 97.35% on a benchmark dataset, beating previous state of the art by more than 27% and closely matching human-level performance
Google DeepMind’s Atari game player outperforms a professional human player in 29 of 49 games tested, trained with over 10 million frames of gameplay
Facebook’s Collaborative Filtering recommends items based on 100 billion ratings, more than a billion users, and millions of items
Google DeepMind’s AlphaGo just won 4-1 against the World Go champion.
www.techarchday.fi
Why now? – Computing power
• Google DeepMind’s AlphaGo runs on 1 920 CPUs and 280 GPUs
– Roughly equivalent to the fastest supercomputer in the world around 2008
– Assuming humans would be able to do 10 games every day, AlphaGo first studied game records for 44 human-years, followed by playing against itself for 8 200 human-years
– Compared to Deep Blue (the famous Chess supercomputer from 1997), uses about 100 000 times more computing power, but still evaluates thousands of times fewer positions meaning that position evaluation in AlphaGo takes hundreds of millions times more computation
– Actually a single GPU wins best amateur professionals!
• What about chess then?– Using the same approaches and only a few GPUs a system can learn to play
chess on a master level in 3 days
www.techarchday.fi
Why now? - Accuracy
• ImageNet Large Scale Visual Recognition Challenge– The standard algorithm benchmark– 1 000 object classes– 1 200 000 training images
www.techarchday.fi
Rapid technology advancements
• Unprecedented data volumes: By 2020, there will be more than 44 zettabytes of data, 35% of which will be considered useful for analysis in 2020
• Decreasing cost of storage: Over the past 30 years, the cost per gigabyte of hard disk data storage has halved every 14 months, from $3.5M in 1980 to $0.03 in 2015
• Virtually unlimited computing power: Public cloud computing will reach almost $70 billion in 2015 worldwide
• Advances in artificial intelligence technologies: AI startups in the US alone have increased 20-fold in the past four years
• Broadening IT scope: 88% of executives agree the IT organization needs to broaden its scope and keep pace with evolving IT needs
Sources: 1“EMC Digital Universe Study,” with data and analysis by IDC, April 2014; 2“Disk Drive Prices (1955-2015)”, John C. McCallum, 2015; 3“ Public Cloud Computing to Reach Nearly $70 billion in 2015 Worldwide, According to IDC,” IDC, Press Release, July 21, 2015; 4”Artificial Intelligence Startups See 302% Funding Jump in 2014,” CB Insights, February 10, 2015; 5 “Calibrating Multi-Speed IT for the Varied Demands of a Multi-Speed Business,” Accenture, 2015.
www.techarchday.fi
Growing number AI vendors
Venture Scanner is tracking 957 AI companies across 13 categories, with a combined funding of $4.8B.
www.techarchday.fi
Virtual Agents or Digital Assistants
Virtual Agents
Virtual Assistants for Personal Computing
Natural Language Question Answering
Virtual Service-Desk Assistant
A new UI: the“new keyboard”
Expert help: a new way to retrieve information
Replaces/helps back-office employees
Siri
Cortana
Silvia
Watson
Wolfram Alpha
IPsoft
NextIT
Nina
Artificial Solutions
Uses Uses
Personal Assistant that uses Google’s Knowledge Graph to answer questions.
EasyAsk’s Quirianswers CRM, ERP & Business Intelligence questions asked in natural language.
Interactive characters with human-like traits and communication styles. Virtual Agents can answer questions and perform business processes.
IPsoft
www.techarchday.fi
Open platforms
• Everyone is in and shares very openly– Apple, Google, Facebook, Microsoft, Baidu, Amazon, IBM,
NVIDIA, Qualcomm, Samsung, Cloudera, Nervana, …
• Data analytics– Storm, Spark, Kafka, Hadoop, HPCC, …
• Machine learning– Scikit-learn, Shogun, Mahout, Samoa, OpenCV, Accord,
Oryx, Weka, Golearn, ConvNetJS, CUDA, …
• Deep learning– TensorFlow, Torch, Caffe, Theano, Chainer, CNTK, mxnet, …
www.techarchday.fi
Machine learning ecosystems
Training as a Service
Vision API
Speech API
Translate API
Pre-trained Services ML-platform as a Service
www.techarchday.fi
Accenture Video Analytics Service Platform
Command & Control Client TerminalsVisualization Operations External Systems
Operational and Predictive Analytics
Core Analytics
Video AnalyticsSocialMediaListening
SensorMonitoring
OperationalSystemsMonitoring
Data Sources
Face Matching License PlateMatching
Crowd Detection
Traffic Monitoring Tracking Flow Management
Anomaly Detection People Counting Vehicle Classification
CCTV Cameras People 3rd Party Sensors3rd Party OperationalSystems
www.techarchday.fi
Video Analytics Applications in Customer Facing Facilities
Retail stores Large scale shopping malls
Facilities operators
Banks / Insurance companies
Public service agencies
Telco agencies
Petrol stations Entertainment parks
Video analytics enable real-time monitoring of facilities and activation of safety and security incident alerts and responses. In addition,
analytical and predictive modeling help optimize facilities operations, and gather marketing insights for commercial tenants
Real-time monitoring
•Visual tracking of people and vehicles, federated solution to monitor multiple buildings
Operational performance measurement
•Service time, wait time, queue monitoring
Commercial space layout optimization
•Identification of hot and cold zones, heat map
Personalized customer experience;
•Dynamic marketing displays, recognition of special customers
Customer analytics
•Demographics analysis, footfall analysis, identification of behavioral trends and anomalies
www.techarchday.fi
Video Analytics Applications in Transportation
Ports Airports
Trains Public Transportation
Video analytics enable real-time monitoring of travel environments and activation of safety and security incident alerts and
responses. In addition, analytical and predictive modeling help optimize operations, and gather marketing insights on
passenger behaviors
Situational awareness
•Integrated central control centers, integration with other sensors and wearables
Onboard and in station real-time monitoring
•Visual tracking of people and vehicles; detection of security and safety incidents
Operational performance measurement
•Service time, wait time, route path analysis
Customer analytics
•Demographics analysis, travelers count, identification of behavioral trends and anomalies
www.techarchday.fi
Classification of car damage level
Given only an image, classify a car as:• Undamaged• Damaged• Totaled
Problem details:
•An Insurance Company wanted to automate claims processing using advanced machine learning technology, namely Deep Learning
•When customers sent a picture of their damaged car, the Company would like to have the ability of automatically detect the level of damage and use it to, for example, order spare parts and possibly detect fraud, if any
•Accenture developed a Convolutional Neural Network algorithm (which belongs to the family of Deep Learning techniques) using a data set of toy images
www.techarchday.fi
About deep learning
Gradient-Based Learning Applied to Document RecognitionLeCun, Bottou, Bengio and Haffner, Proc. of the IEEE 1998
Imagenet Classification with Deep Convolutional Neural NetworksKrizhevsky, Sutskever and Hinton, NIPS 2012
www.techarchday.fi
Damage classification results
90% accuracy achieved
Value delivered:
•By automatically detecting level of damage, an Insurance Company saves on sending a human to assess the damage
•Apply the same technique for more use cases and other lines of businesses like Home Insurance with enhanced complexity and accuracy
•For Auto Insurance, spare parts could be ordered automatically
•For Home Insurance, evaluate building resistance, check if customers are telling the truth about additions to houses, identify multiple damages
•Similarities/differences in damage patterns could be used to detect fraud
www.techarchday.fi
Use case – People counting
• Manual count is based on manual observations of the videos
• Estimation is the number of travellers estimated by the system
• Coloured bounding boxes computed by the different algorithms in place
www.techarchday.fi
Use case – Abandoned objects
• Left object threshold set to 6 seconds
• The passengers nearby are not taken into consideration by the system for this assessment
www.techarchday.fi
AI in enterprise
Text Analytics
Research Assistants
Image Analysis Multimedia Search Cognitive
RoboticsVirtual Agents
Expert Systems
Video Analytics
Identity Analytics
Data Visualization
Domain-specific Calculations
RecommendationSystems
Self-Adjusting IT Systems
Speech Analytics
www.techarchday.fi
Technology vision survey
• 70% of corporate executives say they are making significantly more investments in AI than in 2013
• 55% say they plan to use machine learning and embedded AI solutions extensively
• 78% report the impact will be a significant change or complete transformation over the next three years for their industry
• 84% agree organizations will realize significant competitive advantages with AI
Source: “Accenture Technology Vision Survey,” Accenture, 2016.