JAPAN: ORGANISATION OF PMDA, PHARMACEUTICAL LAWS & REGULATIONS, TYPES OF REGI...
Future of AI & Fabric for Deep Learning (FfDL
1. Future of AI & FfDL
Jim Spohrer (IBM) and Animesh Singh (IBM)
http://slideshare.net/spohrer/intel_20180608_v2
June 8, 2018 - Skype Intel Skype PresentationIntel
Hosts: John Miranda and Michael Jacobson
6/8/2018 IBM #OpenTechAI 1
2. IBM Contacts
6/8/2018 IBM #OpenTechAI 2
Jim Spohrer <spohrer@us.ibm.com>
IBM Research – Almaden
San Jose, CA
Animesh Singh <singhan@us.ibm.com>
IBM Silicon Valley Lab
San Jose, DC
Vijay Bommireddipalli
<vijayrb@us.ibm.com>
CODAIT, San Francisco, CACenter
4. Future of AI
6/8/2018 (c) IBM 2017, Cognitive Opentech Group 4
… when will
your smartphone
be able to take and
pass any online
course? And then
be your coach, so
you can pass too?
7. Every 20 years, compute costs are down
by 1000x
• Cost of Digital Workers
• Moore’s Law can be thought of as
lowering costs by a factor of a…
• Thousand times lower
in 20 years
• Million times lower
in 40 years
• Billion times lower
in 60 years
• Smarter Tools (Terascale)
• Terascale (2017) = $3K
• Terascale (2020) = ~$1K
• Narrow Worker (Petascale)
• Recognition (Fast)
• Petascale (2040) = ~$1K
• Broad Worker (Exascale)
• Reasoning (Slow)
• Exascale (2060) = ~$1K
76/8/2018 (c) IBM 2017, Cognitive Opentech Group
2080204020001960
$1K
$1M
$1B
$1T
206020201980
+/- 10 years
$1
Person Average
Annual Salary
(Living Income)
Super Computer
Cost
Mainframe Cost
Smartphone Cost
T
P
E
T P E
AI Progress on Open Leaderboards
Benchmark Roadmap to solve AI/IA
8. GDP/Employee
6/8/2018 (c) IBM 2017, Cognitive Opentech Group 8
(Source)
Lower compute costs translate into increasing productivity and GDP/employees for nations
Increasing productivity and GDP/employees should translate into wealthier citizens
AI Progress on Open Leaderboards
Benchmark Roadmap to solve AI/IA
9. Leaderboards Framework
AI Progress on Open Leaderboards - Benchmark Roadmap
Perceive World Develop Cognition Build Relationships Fill Roles
Pattern
recognition
Video
understanding
Memory Reasoning Social
interactions
Fluent
conversation
Assistant &
Collaborator
Coach &
Mediator
Speech Actions Declarative Deduction Scripts Speech Acts Tasks Institutions
Chime Thumos SQuAD SAT ROC Story ConvAI
Images Context Episodic Induction Plans Intentions Summarizatio
n
Values
ImageNet VQA DSTC RALI General-AI
Translation Narration Dynamic Abductive Goals Cultures Debate Negotiation
WMT DeepVideo Alexa Prize ICCMA AT
Learning from Labeled Training Data and Searching (Optimization)
Learning by Watching and Reading (Education)
Learning by Doing and being Responsible (Exploration)
2015 2018 2021 2024 2027 2030 2033 2036
6/8/2018 (c) IBM 2017, Cognitive Opentech Group 9
Which experts would be really surprised if it takes less time… and which experts really surprised if it takes longer?
Approx.
Year
Human
Level ->
15. “The best way to predict the future is to inspire the
next generation of students to build it better”
Digital Natives Transportation Water Manufacturing
Energy Construction ICT Retail
Finance Healthcare Education Government
17. Step Comment
GitHub Get an account and read the guide
Learn 3 R's - Read, Redo, Report Read (Medium/arXiv), Redo (GitHub), Report (Jupyter Notebook)
Kaggle Compete in a Kaggle competition
Leaderboards Compete to advance AI progress
Design New Challenges build an AI system that can take and pass any online course, then
switch to tutor-mode and help you pass
Open Source Guide Establish open source culture in your organization
6/8/2018 IBM #OpenTechAI 17
18. Fabric for Deep Learning
FfDL
FfDL Github Page
https://github.com/IBM/FfDL
FfDL dwOpen Page
https://developer.ibm.com/code/open/projects/fabri
c-for-deep-learning-ffdl/
FfDL Announcement Blog
http://developer.ibm.com/code/2018/03/20/fabric-
for-deep-learning
FfDL Technical Architecture Blog
http://developer.ibm.com/code/2018/03/20/democr
atize-ai-with-fabric-for-deep-learning
Deep Learning as a Service within Watson Studio
https://www.ibm.com/cloud/deep-learning
Research paper: “Scalable Multi-Framework
Management of Deep Learning Training Jobs”
http://learningsys.org/nips17/assets/papers/paper_
29.pdf
FfDL
18
https://github.com/IBM/FfDL
19. …that automate
decisions.
…to build models…Use data…
The Enterprise AI Process
19
Gather
Data
Analyze
Data
Machine
Learning
Deep
Learning
Deploy
Model
Maintain
Model
21. Fabric for Deep Learning
https://github.com/IBM/FfDL
FfDL provides a scalable, resilient, and fault
tolerant deep-learning framework
FfDL Github Page
https://github.com/IBM/FfDL
FfDL dwOpen Page
https://developer.ibm.com/code/open/projects/fabri
c-for-deep-learning-ffdl/
FfDL Announcement Blog
http://developer.ibm.com/code/2018/03/20/fabric-
for-deep-learning
FfDL Technical Architecture Blog
http://developer.ibm.com/code/2018/03/20/democr
atize-ai-with-fabric-for-deep-learning
Deep Learning as a Service within Watson Studio
https://www.ibm.com/cloud/deep-learning
Research paper: “Scalable Multi-Framework
Management of Deep Learning Training Jobs”
http://learningsys.org/nips17/assets/papers/paper_
29.pdf
• Fabric for Deep Learning or FfDL (pronounced as ‘fiddle’) is an
open source project which aims at making Deep Learning easily
accessible to the people it matters the most i.e. Data Scientists,
and AI developers.
• FfDL Provides a consistent way to deploy, train and visualize
Deep Learning jobs across multiple frameworks like TensorFlow,
Caffe, PyTorch, Keras etc.
• FfDL is being developed in close collaboration with IBM
Research and IBM Watson. It forms the core of Watson`s Deep
Learning service in open source.
FfDL
21
22. Fabric for Deep Learning
https://github.com/IBM/FfDL
FfDL is built using Microservices architecture
on Kubernetes
• FfDL platform uses a microservices architecture to offer
resilience, scalability, multi-tenancy, and security without
modifying the deep learning frameworks, and with no or minimal
changes to model code.
• FfDL control plane microservices are deployed as pods on
Kubernetes to manage this cluster of GPU- and CPU-enabled
machines effectively
• Tested Platforms: Minikube, IBM Cloud Public, IBM Cloud
Private, GPUs using both Kubernetes feature gate Accelerators
and NVidia device plugins
22
28. And we offer more
Model Asset Exchange
MAX
and
Adversarial Robustness Toolbox
ART
28
29. IBM Model Asset eXchange
MAX
MAX is a one stop exchange to find ML/DL
models created using popular Machine
Learning engines and provides a
standardized approach to consume these
models for training and inferencing.
29
developer.ibm.com/code/exchanges/models/
30. IBM Adversarial Robustness
Toolbox
ART
ART is a library dedicated to adversarial
machine learning. Its purpose is to allow rapid
crafting and analysis of attacks and defense
methods for machine learning models. The
Adversarial Robustness Toolbox provides an
implementation for many state-of-the-art
methods for attacking and defending
classifiers.
30
https://developer.ibm.com/code/open/projects/adver
sarial-robustness-toolbox/
The Adversarial Robustness Toolbox contains
implementations of the following attacks:
Deep Fool (Moosavi-Dezfooli et al., 2015)
Fast Gradient Method (Goodfellow et al., 2014)
Jacobian Saliency Map (Papernot et al., 2016)
Universal Perturbation (Moosavi-Dezfooli et al., 2016)
Virtual Adversarial Method (Moosavi-Dezfooli et al.,
2015)
C&W Attack (Carlini and Wagner, 2016)
NewtonFool (Jang et al., 2017)
The following defense methods are also supported:
Feature squeezing (Xu et al., 2017)
Spatial smoothing (Xu et al., 2017)
Label smoothing (Warde-Farley and Goodfellow, 2016)
Adversarial training (Szegedy et al., 2013)
Virtual adversarial training (Miyato et al., 2017)
32. Model Lifecycle Management
Machine Learning Runtimes Deep Learning Runtimes
Authoring Tools
Cloud Infrastructure as a Service
• Most popular open source frameworks
• IBM best-in-class frameworks
• Create, collaborate, deploy, and monitor
• Best of breed open source & IBM tools
• Code (R, Python or Scala) and no-code/visual
modeling tools
• Fully managed service
• Container-based resource management
• Elastic pay as you go cpu/gpu power
Watson Studio
Tools for supporting the end-to-end AI workflow
33. 3
Train neural
networks in parallel
across NVIDIA
GPUs.
Pay only for what
you use. Auto-
deallocation means
no more
remembering to
shutdown your
cloud training
instances.
Monitor batch training
experiments then
compare cross-model
performance without
worrying about log
transfers and scripts to
visualize results. You
focus on designing your
neural networks. We’ll
manage and track your
assets.
Python client, command
line interface (CLI) or
UI? You choose the
tooling that best fits your
existing workflows.
Training history and
assets are tracked then
automatically transferred
to the customer’s Object
Storage for quick
access.
Deploy models into
production then
monitor them to
evaluate
performance.
Capture new data
for continuous
learning and retrain
models so they
continually adapt to
changing
conditions.
Deep Learning as a Service within Watson Studio
Using FfDL as core
34. Neural Network Modeller within Watson Studio
An intuitive drag-and-drop, no-code interface for designing neural network structure
1950 Nathaniel Rochester (IBM) 701 first commercial computer that did super-human levels of numeric calculations routinely. He worked at MIT on arithmetic unit of WhirlWind I programmable computer.
Dota 2 is most recent August 11, 2017 as a super-human game player in Valve Dota 2 competition – Elon Musk’s OpenAI result.
Miles Bundage tracks gaming progress: http://www.milesbrundage.com/blog-posts/my-ai-forecasts-past-present-and-future-main-post
DOTA2: https://blog.openai.com/more-on-dota-2/
What is beyond Exascale? Zetta (21), Yotta (24)
Time dimension (x-axis) is plus or minus 10 years….
Daniel Pakkala (VTT)
URL: https://aiimpacts.org/preliminary-prices-for-human-level-hardware/
Dan Gruhl:
https://www.washingtonpost.com/archive/business/1983/11/06/in-pursuit-of-the-10-gigaflop-machine/012c995a-2b16-470b-96df-d823c245306e/?utm_term=.d4bde5652826
In 1983 10 GF was ~10 million.
That's 24.55 million in today's dollars.
or 2.4 billion for 1 TF in 1983
Today 1 TF is about $3k http://www.popsci.com/intel-teraflop-chip
The weakest link is what needs to be improved – according to system scientists. Accessing help, service, experts is the weakest link in most systems.
By 2035 the phone may have the power of one human brain – by 2055 the phone may have the power of all human brains.
Before trying to answer the question about which types of sciences are more important – the ones that try to explain the external world or the ones that try to explain the internal world – consider this, slide that shows the different telephones that I have used in my life. I grew up in rural Maine, where we had a party line telephone because we were somewhat remote on our farm in Newburgh, Maine.
However, over the years phones got much better…. So in 2035 or 2055, who are you going to call when you need help?
Where is the variety? Hardware and even software standardizing into modules and algorithms…. Data will standardize next into categories and types…. Experience is where the uniqueness is, and variety and variability, and identity.
By 2036, there will be an accumulation of knowledge as well as a distribution of knowledge in service systems globally. We need to ensure as there is knowledge accumulation that service systems at all scale become more resilient. Leading to the capability of rapid rebuilding of service systems across scales, by T-shaped people who understand how to rapidly rebuild – knowledge has been chunked, modularized, and put into networks that support rapid rebuilding.
Source: Vijay Bommireddipally (CODAIT Director) and Fred Reiss (CODAIT Chief Architect)
URL Amazon: https://www.amazon.com/Knowledge-Rebuild-Civilization-Aftermath-Cataclysm-ebook/dp/B00DMCV5YS/
URL TED Talk: https://www.youtube.com/watch?v=CdTzsbqQyhY
Citation: Dartnell L (2012) The Knowledge: How to Rebuild Civilization in the Aftermath of a Cataclysm. Westminster London: Penguin Books.
Jim Spohrer Blogs:
Grand Challenge: http://service-science.info/archives/2189
Re-readings: http://service-science.info/archives/4416