SKYMIND OVERVIEW - The Channel Company Deep Learning Models with SKIL • Import DL4J, Keras, and...
44
Transcript of SKYMIND OVERVIEW - The Channel Company Deep Learning Models with SKIL • Import DL4J, Keras, and...
Founded 2014Funding $6.3M Clients 14 Enterprise Orgs
3,900 Github Forks300,000+ downloads/mo.
Team 37 employees; 25 engineers; 7 PhDs
SKYMIND OVERVIEW
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
AdamChris
Hacker
House
Skymind circa 2014
OUR USERS
● Cloudera for AI
● A commercially supported enterprise distribution of open-core
software
● We ensure customers succeed in creating custom deep learning
solutions with our product
BUSINESS MODEL
LicensesSkymind sells per core licenses for SKIL.
SupportOngoing model support and maintenance.
TrainingCorporate deep learning workshops led by a Skymind instructor.
WHAT WE SELL
SKIL: Skymind Intelligence Layer
• Workspaces to manage modeling experiments
• Model server to host models produced in workspaces
• Easy REST-based API interface for applications getting predictions from model server
https://skymind.ai/platform
AI: WHY NOW?● Economic, social & technological change make AI both possible & necessary
○ Economic Fifty-two percent of the companies on the Fortune 500 list
have either gone bankrupt, been acquired, ceased to exist, or dropped off
the list since 2000. They need digital transformation, automation and
more accurate predictions. They’ll get those through AI.
○ Social Consumers increasingly accept smart objects such as self-driving
cars and smart appliances infused with voice recognition and CV.
○ Technological A combination of larger data, more powerful hardware
and innovative algorithms have enabled record-breaking accuracy in AI.
● We’re in the midst of a global corporate arms race for AI.
THE WAY BUSINESS MAKES
DECISIONS IS CHANGING
● Real-time Inference
● Evolving Statistical Models
● More Diverse Datasets
FROM BI TO AI
Superhuman accuracy
in machine perception
AI WILL IMPACT EVERY INDUSTRY
THE TOP TECH COMPANIES
ARE POWERED BY AI
● Google (Alphabet)
● Amazon
● Microsoft
Consider Alphabet
● Q2 2017 rev = $26B
● Up 23% YoY (constant currency)
● “Surge in mobile and video ad sales”
● But why are ad sales surging?
THE REST WILL FOLLOW
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
SKYMINDOPEN-CORE AI
CLOUDERA FOR DEEP LEARNING
● An Enterprise Distribution
● Easy Integration with Production Stack
● Supports Major Hardware
● ETL, Training, Inference for DL
TO BUILD AI
YOU NEED 4 THINGS
● Team
● Tools
● Data
● Infrastructure
Team
● Data Scientists/ML specialists
○ Analyze data, prototype models
● Data Engineers
○ Gather, move and store data
● DevOps
○ Maintain AI in production
TOOLS
WHAT DOES ENTERPRISE NEED?
● Open-source (Linux, Hadoop)
● Scalable, Containerized, Fast
● Integrates With Existing Tech (JVM)
● Cross-Team Solution (DevOps, Data Science)
● General-Purpose, Customizable Framework
Operationalizing Deep Learning Models with SKIL
• Import DL4J, Keras, and TensorFlow models natively into the model server
• Hook applications to the model server in the same mindset as you would a JDBC RDBMS application
• Manage, rollback, update models in a way consistent with IT-norms and standards
DATA
● Deep learning needs data to train on
● That data must match the problem you
want to solve
● If you lack labeled data (e.g. face,
name), a labeled data set can be built
● The more, the better
INFRASTRUCTURE
● AI sits on top of the big data stack
● You need software that can gather,
move and store data at scale
● E.g. Hadoop, Spark, Kafka,
Elasticsearch
● And you need a hardware cluster for
compute (GPUs will speed it up.)
Algorithms
WHAT'S AI?
WHAT'S AI?
Algorithms
=
Math & Code
WHAT'S AI?
Algorithms
=
Math & Code
DATA DECISIONS
Human Perception
SENSATION MEANING
Machine Perception
DATA DECISIONS
Prerequisite: Digitization
NUMBERS
BITS
01001101
ANALOG
(REAL LIFE)
WHAT'S AI?
Algorithms
=
Math & Code
DATA
● Images/Video
● Sound/Voice
● Text
● Time Series
DECISIONS
● Classification
○ Name to face
● Clustering
○ Similarity
● Predictions
AI vs. ML vs. DL
AI
AI vs. ML vs. DL
MLDL
AI vs. ML vs. DL● Good old-fashioned AI is based on rules (non-ML AI)
○ Rules tell a computer how to respond to different situations
○ Called expert systems or rules engines
○ Static
● Machine learning
○ ML algorithms adapt when exposed to new data
○ Self-adjusting to improve performance on narrowly defined tasks
○ Dynamic
● Deep learning
○ Computationally intensive
○ Superhuman accuracy
○ State of the art
PRESENT LIMITS
With AI, it can be hard to explain
the difference between
what’s easy and
what’s virtually impossible.
Super-human performance in Go, Texas Hold‘em Image recognition and captioning
Machine language translation Speech recognition and dialog systems36
What AI Is
What AI Isn’t
Strong AI vs. Narrow AI
● Can outperform humans on
every task
● Is embodied
● Has sense of self
● Seeks to maximize
chances of survival,
domination
● Is able to increase its own
intelligence
● Solves one problem well
● Period.
Reinforcement Learning
AlphaGo = DL + RL
Deeplearning4jBuild, train, and deploy neural networks on JVM
RL4JReinforcement learning algorithms on JVM
ND4JHigh performance linear algebra CPU and GPU libraries
SKYMIND TOOLSArbiterHyperparameter search for optimizing neural networks
DataVecData ingestion, normalization, and vectorization
Model ImportImport and deploy neural networks trained in Caffe, Keras, TensorFlow & Theano
Key Skymind Resources
• Platform• https://skymind.ai/platform
• SKIL Documentation• https://docs.skymind.ai/docs
• Blog• https://blog.skymind.ai/
