1 Agenda • Welcome to the cognitive era • Demystify AI / ML / DL • Selected AI Industry Use Cases • How IBM made AI ready for Enterprises ? • Why infrastructure matters ? • The future of AI IBM The Journey Towards Enterprise AI Ahmad El Sayed, Ph.D. Chief Data Scientist Cognitive Systems, IBM, Dubai [email protected]
54
Embed
Agenda IBM - aub.edu.lb · 1 Agenda • Welcome to the cognitive era • Demystify AI / ML / DL • Selected AI Industry Use Cases • How IBM made AI ready for Enterprises ? •
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
1
Agenda
• Welcome to the cognitive era
• Demystify AI / ML / DL
• Selected AI Industry Use Cases
• How IBM made AI ready for Enterprises ?
• Why infrastructure matters ?
• The future of AI
IBM
The Journey TowardsEnterprise AI
Ahmad El Sayed, Ph.D.Chief Data ScientistCognitive Systems, IBM, [email protected]
Welcome to the Cognitive Era
“AI is the fastest growing workload on the planet”, Forrester
190,000
shortage of people with analytical expertise
300%
Increase in AI Spend year over year
$ 320 Billions
The potential impact of AI in the Middle East by 2030
50%
Of CIOs have started or planning to deploy AI solutions
ProblemLow campaign response rates as they are typically executed on mass or segment-based customers.
SolutionBuild predictive model that takes customer 360 view as input and previous campaign responses as output.
Benefits- Increase in campaign response rate- Increase nb of products per customer- Increase in Share of Wallet
Marketing - Next Best Offer
ProblemAs opposed to online channels, brick and mortar retailers lack visibility of customer behavior inside their stores.
SolutionBuild AI models to classify visual features from in-store/mall video, to collect data on consumers profiles, shopping journey, product placement, product touches, and other KPIs
BenefitsOptimize in-store product assortment by analysing profiles and behaviour.Increase of Sales / Margins
Marketing - Footfall Analysis
ProblemHelp doctors to more accurately and more effectively identify the suspected illness areas in the medial images.
SolutionBuild deep learning model to detect, localize and classify suspected disease areas on
medical images, e.g. X-ray, CT, MRI
BenefitsMore accurate disease detectionEfficiency gain to analyse images fasterEarly detection of diseases
Healthcare – Medical Image Analysis
ProblemChallenges of out-of-stock or over-stock result in lost sales and an increase of inventory carrying costs.
SolutionPredict demand based on multimodal data such as historical demand (sales, consumption), marketing data (campaigns), external data (events, weather, demographics, social media)
BenefitsMaximize the service level as well as minimize the inventory cost,increase sales / margins, decrease days of inventory, improve product availability
Operations - Demand Forecasting
Operations - Predictive Maintenance
ProblemReactive and preventive maintenance implies that considerable time/effort is spend on inspecting the wrong asset
SolutionBuild machine learning models to predict failure based on sensor data, asset and maintenance data and then recommend actions to fix part failures.
ProblemManual quality inspection of assets is time-intensive, exhausting, hazardous, inaccurate and sometimes risky.
SolutionBuild AI models to detect, localize and classify defects in batch or real-time on images or videos.
BenefitsImprove inspection accuracyReduce defects and inspection costsEnsure 24/7 operability
Quality Control - Visual Inspection
Security - Worker Safety
ProblemNot enough staff to monitor all cameras placed at dangerous zones,
SolutionBuild AI models to detect workers not respecting security instructions (e.g. no helmet, no vest)
BenefitsReduce risk of incidents at dangerous zones without having to hire more staff to monitor live CCTV cameras.
ProblemNot enough security staff to monitor all cameras at all time. VMS aren’t flexible to detect new patterns, large number of false alerts, static patterns.
SolutionBuild AI modes to detect suspicious objects or activities and generate alerts in real-time to prevent crimes.
BenefitsDecrease in false and missed alertsDecrease in average search timeDecrease in number of crimes
Security - Video Surveillance
ProblemFraudsters are constantly finding new schemes to fraud the system, so it’s important to have multi-channel monitoring
SolutionBuild AI models that can constantly learn to detect evolving thefts techniques from structured and unstructured data
BenefitsReduce fraud lossesPrevent reputational damagePrevent fraud in real-time
Security - Fraud Detection
ProblemUsers spend 2 weeks in average to watch a program and extract the top scenes in 2 minutes trailer, which is very time-consuming
SolutionAI models are built to analyze the video and audio to rank scenes by the level of excitement and then select the top ones
BenefitsProducing a trailer with our models take now 2 hours as opposed to 2 weeks which free stafffor more quality tasks.
Entertainment - Trailer Production
How IBM made AI ready for the Enterprise ?
32
Enable non-Data Scientists to use AI(PowerAI Vision & Others)
Higher Productivity for Data Scientists(Faster Training with Larger Models)
Integrated & Supported AI Platform
Caffe
IBM PowerAI – the Enterprise Offering for Deep Learning
GPU-Accelerated
Power Servers
Storage
PowerAI Base on IBM Power AC922
33
• Co-Optimized Software + Hardware
• Enterprise Software Distribution
• Best Server for Enterprise AI with super accelerated highways between CPU-GPU and GPU-GPU
• Performance optimized for large model support and distributed deep learning
• Enterprise Support L1-L3
Caffe
34
34
34
3.1 Hours
49 Mins
0
2000
4000
6000
8000
10000
12000
Xeon x86 2640v4 w/ 4xV100 GPUs
Power AC922 w/ 4x V100GPUs
Tim
e (s
ecs)
Caffe with LMS (Large Model Support)Runtime of 1000 Iterations
3.8x Faster
GoogleNet model on Enlarged ImageNet Dataset (2240x2240)
34
1 System 64 Systems
16 Days Down to 7 Hours58x Faster
16 Days
7 Hours
Near Ideal Scaling to 256 GPUs
ResNet-101, ImageNet-22K
1
2
4
8
16
32
64
128
256
4 16 64 256
Spee
du
p
Number of GPUs
Ideal Scaling
DDL Actual Scaling
95%Scaling
with 256 GPUS
Caffe with PowerAI DDL, Running on Minsky (S822Lc) Power System