THE WAY BUSINESS MAKES
DECISIONS IS CHANGING
● Real-time Inference
● Evolving Statistical Models
● More Diverse Datasets
A SHIFT FROM BIG DATA
TO MACHINE INTELLIGENCE
Superhuman accuracy
with machine perception
AI WILL IMPACT EVERY INDUSTRY
THE TOP TECH COMPANIES
ARE POWERED BY DL
● Google
● Facebook
● Amazon
● Microsoft
THE REST WILL FOLLOW ... OR DIE
WHAT ARE THE REQUIREMENTS
FOR ENTERPRISE AI?
● Open-source (Linux, Hadoop)
● Scalable, Containerized, Fast
● Integrates With Existing Tech (JVM)
● Cross-Team Solution (DevOps, Data Science)
● General-Purpose, Customizable Framework
SKYMIND GIVES
BIG COMPANIES
DEEP LEARNING
AN OPEN-CORE COMPANY
CLOUDERA FOR AI
● Enterprise Distribution
● Easy Integration with Production Stack
● Supports Major Hardware
● ETL, Training, Inference for DL
AI-DRIVEN SOLUTIONS FOR
● BANKS
● TELCOS
● HEALTHCARE
● RETAIL
● MANUFACTURING
● DATA CENTERS
USE CASES
● FRAUD/ANOMALY DETECTION
○ BILLING, TELCO, ID
● IMAGE RECOGNITION
● PREDICTIVE ANALYTICS
○ MARKET FORECASTING
○ CHURN/TURNOVER
○ HARDWARE BREAKDOWNS
PRODUCT
WE BUILT
DEEPLEARNING4J
THE MOST WIDELY USED AI
FRAMEWORK FOR JAVA
THE BRIDGE JOINING
THE JVM PRODUCTION STACK
& ACCELERATED HARDWARE
SPARK
JAVACPP
FPGAARM
Existing Planned
KEY INTEGRATIONS: SPARK, MESOS, KAFKA & HADOOP
FULL AI STACK: THE SKYMIND INTELLIGENCE LAYER (SKIL)
SKYMIND UI FOR DEPLOYMENT ON MESOS & DC/OS
● ON PREMISE
● PUBLIC CLOUD
○ AWS
○ AZURE
○ GCE
● HYBRID CLOUD
● DESKTOP
○ WINDOWS
○ MAC
● ANDROID
DOCKER
CONTAINERIZED DEEP LEARNING FOR ANY PLATFORM
DC/OS + SPARK
MESOS
DeepLearning4j TheanoTensorFlow
Intuitive Python API: Keras
Multi-GPUS Single GPU
Java API Scala API
JVM Big Data Stack:
Hadoop, Spark,
Kafka, ElasticSearch
CaffeTorch
Lua API
Image Only
Commercial Support
Platform Neutral: On-prem/AWS..
Optimized for
Google Cloud
& TPUs
General Purpose Deep Learning Platforms: For Image, Video, Sound, Text, Time Series Data
Python API
Multi-GPUS
Enterprise distro certified on
CDH, HDP, Kerberos
MxNet
R, Python, etc.
MAJOR DEEP LEARNING LIBRARIES COMPARED
MAJOR DEEP LEARNING LIBRARIES COMPARED
ROADMAP
● INTERPRETABLE MODELS
● MORE PRE-TRAINED MODELS
● VERTICAL SPECIFIC GUIs
● FEATURE PARITY WITH KERAS
PROOF
100,000+ MONTHLY DOWNLOADS
TRACTION
TRAFFIC
• 160,000+ Web site Hits per Month
• SEO: No. 1 “deep learning java”, “deep learning
open source” …
DOWNLOADS
• 100,000+ downloads/mo
• 20% mo/mo growth in downloads
PUBLICITY
• Wired, TechCrunch, WSJ, Bloomberg, VB
TRACTION 2
Skymind Projects on Github
• 5,000 stars; 2000+ forks
• Top 0.001% of all Java projects
• Comparisons
• Mesos: ~2,700 stars
• Reported $1B valuation
• Impala : ~1,700 stars
• (Supported by Cloudera,
a $4.1B enterprise co.
with $1.2B in funding…)
GO-TO-MARKET PARTNERS
• Nvidia: Co-promotion, joint customers, Inception
• Intel: Deep-Learning Partner (MOU)
• Cloudera: Certified, co-promotion
• Hortonworks: Certified, DL4J apps built
• IBM: Customer. Multiple teams, uses. Watson IoT
• Lightbend (Lagom for micro-services)
KEY OPEN-SOURCE USERS
• Cornell
• Stanford
• JPMorgan
• IBM
• U. Massachusetts
• Yale
• Harvard
• MIT
OUR BOOK
INVESTORS
• SV ANGEL (RON CONWAY)
• GPV (ORACLE CEO RAY LANE)
• Y COMBINATOR
• TENCENT
• MANDRA CAPITAL
• RISING TIDE FUND
• LIQUID 2 VENTURES
• AMIT SINGHAL (GOOGLE SEARCH)
• KRISHNA BHARAT (GOOG. NEWS)
• KEVIN MAHAFFEY (LOOKOUT)
TEAM
FOUNDERS (YC W16)
Deep learning @GalvanizeU
• Author: O’Reilly’s “Deep learning: A
Practitioner’s Guide” Mar. 2016
• Speaker: Hadoop Summit, OSCon, Tech
Planet, GigaOM
• 3x startup founder
• CS/Biz @Michigan Tech
ADAM GIBSON, CTO CHRIS NICHOLSON, CEO
Sequoia’s FutureAdvisor
• As a recruiter: Helped triple team
through Series B to 45 staff
• As PR: Helped drive 45x rev. and AUM growth
($650M in June 2015)
• New York Times correspondent covering tech,
M&A: 2006-2011
SKYMIND ENGINEERS
FRANCOIS GARILLOT
• Deep learning Systems Engineer
• Former Lightbend, Swisscom
• PhD in C.S. from l'Ecole Polytechnique
JOSH PATTERSON
• Head Field Engineer
• ex-Principal Architect Cloudera
• O’Reilly Co-Author
SAMUEL AUDET
• Author of JavaCPP (Cython for Java)
• PhD in Computer Vision from TIT
SUSAN ERALY
• Sr. ASIC Engineer at NVIDIA
• Physical Design Eng. H-P
ALEX BLACK
• Doctoral Candidate (math)
• Java Software engineer
V. KOKORIN
• NLP solutions builder
• GPU optimization
CASE STUDIES
TELECOM CASE STUDY
FRANCE TELECOM'S MOBILE UNIT
ORANGE
PROBLEM
Orange has a fraud problem costing it tens of millions of dollars a
year. It's called bypass fraud: bad actors on Orange's network avoid
paying international calling fees while still routing calls from one
country to another through VOIP. Orange's previous solution
needed days to detect a third-party service cheating on fees. In that
time, the actor could make tremendous profits. Orange's previous
system involved a combination of hard rules and SQL queries over
data stored on a Hadoop cluster.
SOLUTION
Skymind's neural nets are able to detect anomalous calling patterns
that locate and identify bypass fraud with just a few hours' data.
Using Spark on top of Orange's Hadoop cluster, Skymind trained a
neural network architecture to detect unusual behavior that was
escalated to Orange's human analyst team, who then decide which
users to shut down. The results obtained are better than any
previous solution produced. The solution is now deployed on
Orange's servers in France.
BANK CASE STUDY
LARGE ASIAN BANK
PROBLEM
A global bank processes payments from around the world and
needs to detect unusual behavior such as fraud and money
laundering. The bank had built its own solution which it maintained
and extended until it discovered Deeplearning4j. That solution was
brittle and rules based, and had trouble evolving quickly in an
adversarial situation. The process of developing that in-house
framework required the attention of rare feature engineers and
produced mediocre results.
SOLUTION
Skymind's enterprise distribution of Deeplearning4j gave the bank
modular, well-maintained and hardware-optimized code which it
could extend to surface anomalous behavior. These filters were
integrated with data pipelines that included Kafka, Spark and HDFS,
and run on top of ND4J, a scientific computing library Skymind built,
which bridges academic hardware acceleration on GPUs with the
big data ecosystem of the JVM. The bank's engineers no longer had
to manually code their rules and filters, and were able to extend the
bank's infrastructure in other ways.
DATA CENTER CASE
STUDY
CANONICAL
PROBLEM
Enterprise clients pay Canonical to manage instances of
OpenStack, an open-source software platform deployed as
infrastructure as a service. OpenStack runs on top of commodity
hardware, and Canonical is responsible for its security and smooth
operation.
SOLUTION
Skymind built DeepStack, a nervous system for OpenStack.
DeepStack detects:
● Network intrusions through data packet inspections
● Imminent hardware breakdowns by analyzing server logs
DeepStack is used for cybersecurity and to rebalance workloads.
IMAGE RECOGNITION
CASE STUDY
BERNIE.AI
PROBLEM
Bernie.ai is an innovative dating app and personal assistant. Bernie
needed a method to quickly analyze and cluster photographs of
faces in order to offer its users attractive potential dates based on
their likes and dislikes. More accurate recommendations reduce
dating fatigue. Existing deep-learning tools including well-known
Python frameworks Bernie experimented with were too slow.
SOLUTION
Skymind's fast, accurate image recognition algorithms gave Bernie
a speed improvement of 3,700% and accuracy of more than 98% in
facial recognition.
help@skymind.io

Chris Nicholson, CEO Skymind at The AI Conference

  • 2.
    THE WAY BUSINESSMAKES DECISIONS IS CHANGING ● Real-time Inference ● Evolving Statistical Models ● More Diverse Datasets
  • 3.
    A SHIFT FROMBIG DATA TO MACHINE INTELLIGENCE Superhuman accuracy with machine perception
  • 4.
    AI WILL IMPACTEVERY INDUSTRY
  • 5.
    THE TOP TECHCOMPANIES ARE POWERED BY DL ● Google ● Facebook ● Amazon ● Microsoft
  • 6.
    THE REST WILLFOLLOW ... OR DIE
  • 7.
    WHAT ARE THEREQUIREMENTS FOR ENTERPRISE AI? ● Open-source (Linux, Hadoop) ● Scalable, Containerized, Fast ● Integrates With Existing Tech (JVM) ● Cross-Team Solution (DevOps, Data Science) ● General-Purpose, Customizable Framework
  • 8.
  • 9.
    AN OPEN-CORE COMPANY CLOUDERAFOR AI ● Enterprise Distribution ● Easy Integration with Production Stack ● Supports Major Hardware ● ETL, Training, Inference for DL
  • 10.
    AI-DRIVEN SOLUTIONS FOR ●BANKS ● TELCOS ● HEALTHCARE ● RETAIL ● MANUFACTURING ● DATA CENTERS
  • 11.
    USE CASES ● FRAUD/ANOMALYDETECTION ○ BILLING, TELCO, ID ● IMAGE RECOGNITION ● PREDICTIVE ANALYTICS ○ MARKET FORECASTING ○ CHURN/TURNOVER ○ HARDWARE BREAKDOWNS
  • 12.
  • 13.
    WE BUILT DEEPLEARNING4J THE MOSTWIDELY USED AI FRAMEWORK FOR JAVA
  • 14.
    THE BRIDGE JOINING THEJVM PRODUCTION STACK & ACCELERATED HARDWARE SPARK JAVACPP FPGAARM Existing Planned
  • 15.
    KEY INTEGRATIONS: SPARK,MESOS, KAFKA & HADOOP
  • 16.
    FULL AI STACK:THE SKYMIND INTELLIGENCE LAYER (SKIL)
  • 19.
    SKYMIND UI FORDEPLOYMENT ON MESOS & DC/OS
  • 20.
    ● ON PREMISE ●PUBLIC CLOUD ○ AWS ○ AZURE ○ GCE ● HYBRID CLOUD ● DESKTOP ○ WINDOWS ○ MAC ● ANDROID DOCKER CONTAINERIZED DEEP LEARNING FOR ANY PLATFORM DC/OS + SPARK MESOS
  • 22.
    DeepLearning4j TheanoTensorFlow Intuitive PythonAPI: Keras Multi-GPUS Single GPU Java API Scala API JVM Big Data Stack: Hadoop, Spark, Kafka, ElasticSearch CaffeTorch Lua API Image Only Commercial Support Platform Neutral: On-prem/AWS.. Optimized for Google Cloud & TPUs General Purpose Deep Learning Platforms: For Image, Video, Sound, Text, Time Series Data Python API Multi-GPUS Enterprise distro certified on CDH, HDP, Kerberos MxNet R, Python, etc. MAJOR DEEP LEARNING LIBRARIES COMPARED
  • 23.
    MAJOR DEEP LEARNINGLIBRARIES COMPARED
  • 24.
    ROADMAP ● INTERPRETABLE MODELS ●MORE PRE-TRAINED MODELS ● VERTICAL SPECIFIC GUIs ● FEATURE PARITY WITH KERAS
  • 25.
  • 26.
  • 27.
    TRACTION TRAFFIC • 160,000+ Website Hits per Month • SEO: No. 1 “deep learning java”, “deep learning open source” … DOWNLOADS • 100,000+ downloads/mo • 20% mo/mo growth in downloads PUBLICITY • Wired, TechCrunch, WSJ, Bloomberg, VB
  • 28.
    TRACTION 2 Skymind Projectson Github • 5,000 stars; 2000+ forks • Top 0.001% of all Java projects • Comparisons • Mesos: ~2,700 stars • Reported $1B valuation • Impala : ~1,700 stars • (Supported by Cloudera, a $4.1B enterprise co. with $1.2B in funding…)
  • 29.
    GO-TO-MARKET PARTNERS • Nvidia:Co-promotion, joint customers, Inception • Intel: Deep-Learning Partner (MOU) • Cloudera: Certified, co-promotion • Hortonworks: Certified, DL4J apps built • IBM: Customer. Multiple teams, uses. Watson IoT • Lightbend (Lagom for micro-services)
  • 30.
    KEY OPEN-SOURCE USERS •Cornell • Stanford • JPMorgan • IBM • U. Massachusetts • Yale • Harvard • MIT
  • 31.
  • 32.
    INVESTORS • SV ANGEL(RON CONWAY) • GPV (ORACLE CEO RAY LANE) • Y COMBINATOR • TENCENT • MANDRA CAPITAL • RISING TIDE FUND • LIQUID 2 VENTURES • AMIT SINGHAL (GOOGLE SEARCH) • KRISHNA BHARAT (GOOG. NEWS) • KEVIN MAHAFFEY (LOOKOUT)
  • 33.
  • 34.
    FOUNDERS (YC W16) Deeplearning @GalvanizeU • Author: O’Reilly’s “Deep learning: A Practitioner’s Guide” Mar. 2016 • Speaker: Hadoop Summit, OSCon, Tech Planet, GigaOM • 3x startup founder • CS/Biz @Michigan Tech ADAM GIBSON, CTO CHRIS NICHOLSON, CEO Sequoia’s FutureAdvisor • As a recruiter: Helped triple team through Series B to 45 staff • As PR: Helped drive 45x rev. and AUM growth ($650M in June 2015) • New York Times correspondent covering tech, M&A: 2006-2011
  • 35.
    SKYMIND ENGINEERS FRANCOIS GARILLOT •Deep learning Systems Engineer • Former Lightbend, Swisscom • PhD in C.S. from l'Ecole Polytechnique JOSH PATTERSON • Head Field Engineer • ex-Principal Architect Cloudera • O’Reilly Co-Author SAMUEL AUDET • Author of JavaCPP (Cython for Java) • PhD in Computer Vision from TIT SUSAN ERALY • Sr. ASIC Engineer at NVIDIA • Physical Design Eng. H-P ALEX BLACK • Doctoral Candidate (math) • Java Software engineer V. KOKORIN • NLP solutions builder • GPU optimization
  • 36.
  • 37.
    TELECOM CASE STUDY FRANCETELECOM'S MOBILE UNIT ORANGE
  • 38.
    PROBLEM Orange has afraud problem costing it tens of millions of dollars a year. It's called bypass fraud: bad actors on Orange's network avoid paying international calling fees while still routing calls from one country to another through VOIP. Orange's previous solution needed days to detect a third-party service cheating on fees. In that time, the actor could make tremendous profits. Orange's previous system involved a combination of hard rules and SQL queries over data stored on a Hadoop cluster.
  • 39.
    SOLUTION Skymind's neural netsare able to detect anomalous calling patterns that locate and identify bypass fraud with just a few hours' data. Using Spark on top of Orange's Hadoop cluster, Skymind trained a neural network architecture to detect unusual behavior that was escalated to Orange's human analyst team, who then decide which users to shut down. The results obtained are better than any previous solution produced. The solution is now deployed on Orange's servers in France.
  • 40.
  • 41.
    PROBLEM A global bankprocesses payments from around the world and needs to detect unusual behavior such as fraud and money laundering. The bank had built its own solution which it maintained and extended until it discovered Deeplearning4j. That solution was brittle and rules based, and had trouble evolving quickly in an adversarial situation. The process of developing that in-house framework required the attention of rare feature engineers and produced mediocre results.
  • 42.
    SOLUTION Skymind's enterprise distributionof Deeplearning4j gave the bank modular, well-maintained and hardware-optimized code which it could extend to surface anomalous behavior. These filters were integrated with data pipelines that included Kafka, Spark and HDFS, and run on top of ND4J, a scientific computing library Skymind built, which bridges academic hardware acceleration on GPUs with the big data ecosystem of the JVM. The bank's engineers no longer had to manually code their rules and filters, and were able to extend the bank's infrastructure in other ways.
  • 43.
  • 44.
    PROBLEM Enterprise clients payCanonical to manage instances of OpenStack, an open-source software platform deployed as infrastructure as a service. OpenStack runs on top of commodity hardware, and Canonical is responsible for its security and smooth operation.
  • 45.
    SOLUTION Skymind built DeepStack,a nervous system for OpenStack. DeepStack detects: ● Network intrusions through data packet inspections ● Imminent hardware breakdowns by analyzing server logs DeepStack is used for cybersecurity and to rebalance workloads.
  • 47.
  • 48.
    PROBLEM Bernie.ai is aninnovative dating app and personal assistant. Bernie needed a method to quickly analyze and cluster photographs of faces in order to offer its users attractive potential dates based on their likes and dislikes. More accurate recommendations reduce dating fatigue. Existing deep-learning tools including well-known Python frameworks Bernie experimented with were too slow.
  • 49.
    SOLUTION Skymind's fast, accurateimage recognition algorithms gave Bernie a speed improvement of 3,700% and accuracy of more than 98% in facial recognition.
  • 51.