London, May 2019 Moty Fania Principle Engineer (CTO), Advanced Analytics Sales
London, May 2019
Moty FaniaPrinciple Engineer (CTO), Advanced Analytics
Sales
Legal notices
This presentation is for informational purposes only. INTEL MAKES NO WARRANTIES, EXPRESS OR IMPLIED, IN THIS SUMMARY.
For more complete information about performance and benchmark results, visit www.intel.com/benchmarks
Intel and the Intel logo are trademarks of Intel Corporation in the U.S. and/or other countries.
* Other names and brands may be claimed as the property of others.
Copyright © 2019, Intel Corporation. All rights reserved.
2
3
Descriptive Analytics
Diagnostic Analytics
Predictive Analytics
Prescriptive Analytics
Cognitive Analytics
Operational Analytics
AdvancedAnalytics
ForesightWhat Will Happen,
When, and Why
HindsightWhat Happened
Self-Learning and Automation
Computerized human thought and actions
InsightWhat Happened
and Why
Va
lue
Difficulty
Simulation Driven Analysis and
Decision Making
10
Harnessing Analytics
Narrow AI
Also referred to as “weak AI.” This is AI that works within a very limited context, and can’t take on tasks beyond its field. It is the only form of Artificial Intelligence achieved so far.
Artificial intelligence is about replacing human decision making
▪ These are not repetitive tasks, but rather judgment-based decisions
▪ Measured by the quality of decisions & actions taken to achieve a desired objective
▪ Machine learning as the key technology to use a variety of inputs , recognize patterns, predict future outcomes and make decisions.
4
AI in a Nutshell
What is
AI?
About Us – Advanced Analytics @ IntelHo
wVa
lue
Embed Learning HW Validation Product Dev Sales Industrial IOT Health
Improve products power/
performance
Cut product time to market
Increaserevenue
Reduce Manufacturing
cost
Improve clinical trials outcome
Reduce test cost and improve
quality
Adaptive & personalized
HW
Automated validation with
Context
Autonomous accounts coverage
Proactive actuation to
changes
Continuance Monitoring at
home
Personalized units testing
-----------------
5
Vision : Put AI to work for human experts
Why AI Platforms?
• Developing an excellent AI model(s) that solve a business problem is essential but not sufficient
• Successful AI implementation requires deep integration into business processes and the ability to continuously and rapidly deploy AI models to the production environment of a biz domain
Vertical Specific AI Platforms
Infrastructure (on premise or Public Cloud)
Platforms / SW “Building Blocks”
• Dedicated, Specialized AI platforms are built using standard, open source building blocks to enable this continuous, closed feedback loop and offer better TTM and lower TCO
SalesHealthHW Validation
-----------------
Product Dev Industrial AI
The Characteristics of AI Platforms
Key Characteristics:
• Deep integration to vertical's biz processes
• Continuous delivery / deployment
• On-Prem, Cloud, Hybrid• Closed feedback loop &
Actuation
• Designed for experts• No GUI • Code based• Open Source
AI Platforms enable a self-sustained, on-going AI Service in production with built-in integration points for different kinds of models
Sales AI
Sense Reason Interact Learn
Information Action Results
Real time scan of customer information
Detect intent to buy and potential opportunities
Automate a personalized best
action
Automatically learn based on customer
response
Sales AI Platform – What is required?
Online streaming
system
Agile Micro-services,
architecture
Specialized data stores –
Knowledge Graph, Search, Big data
object store
NLU and auto labeling
Real time DL inference
service
Scalable computation Engine
Online streaming system
TwitterService Message Bus
SQLData Stores
CrawlService
Seed Queue
Handles Queue
Reasoning Agents
* Intel’s Sales AI uses APIs provided by Twitter to collect data only from accounts the user has designated as public and adheres to individual websites’ terms of use. The only data collected is what is permitted by the site owner under their terms of use agreement.
Micro-services, architecture
Message Bus - Broker
Information extraction …
Interactive Agents
Inference Agent
GraphBuilder
Flexible Data Store(s)
Search(Engine)
SQL Big data Store Knowledge Graph
Why Knowledge Graphs?
• We are surrounded by entities , which are connected by relations
• Graphs are a natural way to represent entities and their relationships
• We need a structured and formal representation of knowledge over time
• Graphs can be managed efficiently
• Enable reasoning to infer new knowledge
Works forCompany B
CompanyA
CompanyC
Person
David Smith
Acquired
Partners with
ProductX
Buys
Semantic descriptions of entities and their relationships
14
• Microservice architecture
• Asynchronous stream processing using Kafka as the message bus
• Easily deployable with Docker, K8S and Helm
• Configuration driven - configurable schema vs. code
• Decoupled from specific Graph technology
• Support incremental updates + ongoing refresh process
“Graphier” – A service for building and sustaining graphs
Message Bus
Asynchronous Graph Transformers
• Always ON (Docker containers)
• Stateless
• Configuration driven behavior
• Each handles a certain graph entity
• May produce of work for others!
• Implemented with Kafka streams
15
Consume from Work Queue
Extract Data
Convert to Graph Semantic
Produce to Broker
* Other names and brands may be claimed as the property of others.
• Transformer – Transforms raw string data from message bus to specific graph semantics
• Extractor - Implements the logic for a specific entity extraction from an external source
16
API
GraphBuilderTransformers ExtractorsData Loaders
Data Refresh
Message Bus
High level architecture
* Other names and brands may be claimed as the property of others.
Externalsources
17
GraphBuilder
“Company”Transformer
“Company”Extractor
Graph Builder Queue
Company Queue
Sample Flow
* Other names and brands may be claimed as the property of others.
CRM Companies Data
Data Loader
{Company: ABC,Country: US
}
{Nodes: [],Edges: []
}
Pro
du
ce
Co
nsu
me
“Product”Extractor
Company Extract Queue
Product Queue
{Product: X,
}
{Nodes: [],Edges: []
}
Pro
du
ce
{Nodes: [],Edges: []
}
18
Semantic Data Converter
• Enables adding new entities’ definitions without code changes
• Converts Business entities into Graph semantics
• YAML configurable
Data Converter{
BizName: ABC,Country: US
}
{Nodes: [],Edges: []
}
Company
Implemented as a generic module, consumed by Transformers and Extractors
Deep Learning Inference System
Inference Service
… ……… … ………
… ……… … ………
ML1 ML1
Classification Combiner
1
2
…
n
Cache
Request(POST)
Predict (GET)
• Many AI capabilities involve usage of Deep Learning
• NLP / NLU
• Visual inspection
• Recommendation Systems
• Implemented production grade, managed, inference system that serves DL models and enables online, closed feedback loop
Archelon
Reduces TCO and TTM
• Production Inference service for DL models
• Smart in-memory cache for data batching , sequencing
• Fast, scalable APIs for data ingestion & real time responses
• Full Scalability
Natural Language Understanding
• Named-entity recognition (NER)
• Text-classification M&A
• Relations extraction Amazon To Acquire Ring
• Intent classification Amazon Enters Industry Home Security
• Sentiment analysis Positive
Many (text-related) business problems require supervised methods
Snorkel is an open-source framework capable of training models for supervised problems without manual-labeling
Helps avoid rule-based classification and Amazon Mechanical Turk for labeling
We’re excited to report that Amazon buys Ring to get into home security
business.
RAY – A Scalable computation engine
• Ray is a flexible, high-performance, distributed execution framework for Emerging AI Applications (UC Berkeley)
• A scalable system for parallel and distributed Python.
• Shortens path to production (scalability, throughput, tolerance)
• Dev & DS friendly – Powerful yet easy actor system
• Superior performance to Spark in certain use cases
Ray on standalone mode is on an average 3x faster and on cluster mode is on an average 18x faster than the scenarios performed without Ray
22
Putting it all together
Consumer
Consumer
Consumer
Databases
Graph Builder
Relations analytics
Training
Action service - Points
Message Bus
Crawler
Deep LearningInference
NLU / Information extraction
API
Computation Engine
Action service - Training
Action service – Sales Assists
Action service Matchmaking
…
Archelon RAY
news
* Intel’s Sales AI uses APIs provided by Twitter to collect data only from accounts the user has designated as public and adheres to individual websites’ terms of use. The only data collected is what is permitted by the site owner under their terms of use agreement.
CRM
Sales Assist User Interface
Sales Assist provides relevant customer activity to the account manager with details that provide additional information.
Sales Assist User Interface – Assist Detail
With the “Details” for a “Viewed Intel product” assist, the account manager can see the specific pages the customer visited on intel.com and Sales Assist highlights that the customer has not previously purchased the product.
An Autonomous sales Example
Sense direct Intent
Intel.
com
Train
ing
Customer DataSINGLE
CUSTOMER VIEW
Recommender System
Reasoning Interact Learn
Communication
in 15 different languages
~150K
Contacts
• Open rate• Recommended
SKUs purchased in relevant period
A/B Testing
Sent to ~3% of contacts
IT@INTEL
Learn more about Intel IT’s initiatives at: www.intel.com/IT26
Sharing Intel IT Best Practices With the World
Copy protected with Online-PDF-No-Copy.com