Nike Tech Talk, Portland, 2017-08-10
https://niketechtalks-aug2017.splashthat.com/
O'Reilly Media gets to see the forefront of trends in artificial intelligence: what the leading teams are working on, which use cases are getting the most traction, previews of advances before they get announced on stage. Through conferences, publishing, and training programs, we've been assembling resources for anyone who wants to learn. An excellent recent example: Generative Adversarial Networks for Beginners, by Jon Bruner.
This talk covers current trends in AI, industry use cases, and recent highlights from the AI Conf series presented by O'Reilly and Intel, plus related materials from Safari learning platform, Strata Data, Data Show, and the upcoming JupyterCon.
Along with reporting, we're leveraging AI in Media. This talk dives into O'Reilly uses of deep learning -- combined with ontology, graph algorithms, probabilistic data structures, and even some evolutionary software -- to help editors and customers alike accomplish more of what they need to do.
In particular, we'll show two open source projects in Python from O'Reilly's AI team:
• pytextrank built atop spaCy, NetworkX, datasketch, providing graph algorithms for advanced NLP and text analytics
• nbtransom leveraging Project Jupyter for a human-in-the-loop design pattern approach to AI work: people and machines collaborating on content annotation
3. Research
questions:
▪ How
do
we
personalize
learning
experiences,
across
ebooks,
videos,
conferences,
computable
content,
live
online
courses,
case
studies,
expert
AMAs,
etc.
▪ How
do
we
help
experts
—
by
definition,
really
busy
people
—
share
knowledge
with
their
peers
in
industry?
▪ How
do
we
manage
the
role
of
editors
at
human
scale,
while
technology
and
delivery
media
evolve
rapidly?
▪ How
do
we
help
organizations
learn
and
transform
continuously?
▪ Can
we
accomplish
these
goals
by
leveraging
AI
in
Media?
3
7. AI
is
real,
but
why
now?
▪ Big
Data:
machine
data
(1997-‐ish)
▪ Big
Compute:
cloud
computing
(2006-‐ish)
▪ Big
Models:
deep
learning
(2009-‐ish)
The
confluence
of
factors
created
a
business
environment
where
AI
could
become
mainstream
AR/VR
combined
with
embedded
computing
and
reinforcement
learning
may
bring
it
to
a
next
level
7
8. Benchmark:
achieving
human
parity
2016-‐10-‐12:
Microsoft
researchers
reach
human
parity
in
conversational
speech
recognition
Achieving
Human
Parity
in
Conversational
Speech
Recognition
W.
Xiong,
et
al.
Microsoft
8
9. Big
picture
▪ The
current
state
of
machine
intelligence
3.0
Shivon
Zilis,
James
Cham
Bloomberg
Beta
(annual
landscape)
▪ The
Future
of
Machine
Intelligence
David
Beyer
Amplify
Partners
(report)
▪ Artificial
Intelligence:
Teaching
Machines
to
Think
Like
People
Jack
Clark
Open
AI
(report)
▪ The
AI
Conf
O’Reilly
Media
and
Intel
partnership
(industry
conference)
9
11. “Consider
the
shift
from
steam
to
electric
power:
it
took
a
generation
before
factory
managers
understood
they
could
reconfigure
the
physical
arrangement.
AI
may
be
quicker
adoption,
but
faces
similar
extremes
of
cognitive
embrace.”
–
David
Beyer
Amplify
Partners
11
12. Immediate
impact
of
AI
12
personal
op-‐ed:
the
combination
of
advances
with
UX,
DevOps,
AI
together
–
specifically
–
is
taking
off
the
table
some
previous
needs
for
what
we’d
called
“software
engineering”
–
which
must
now
undergo
major
changes
14. 2017
highlights
from
leading
teams
▪ TensorFlow:
Machine
learning
for
everyone
Rajat
Monga
Google
▪ Distributed
deep
learning
on
AWS
using
MXNet
Anima
Anandkumar
Amazon
▪ Squeezing
deep
learning
onto
mobile
phones
Anirudh
Koul
Microsoft
14
22. 22
Current
themes
among
leading
AI
teams:
▪ scale
up
to
solve
complex
problems
(big
models)
▪ optimize
to
deploy
consumer
products
(low
power)
Trending
strategy…
23. 23
Most
popular
content,
among
thousands
of
enterprise
organizations:
Hands-‐On
Machine
Learning
with
scikit-‐learn
and
TensorFlow
Aurélien
Géron
Python
FTW.
Along
with
Keras,
PyTorch,
Caffe,
etc.
Trending
methods…
24. UC
Berkeley
RISELab
24
▪ https://rise.cs.berkeley.edu/
▪ enable
machines
to
take
rapid,
intelligent
actions
based
on
real-‐time
data
and
context
from
the
world
around
them
▪ shift
away
from
prior
emphasis
on
JVM-‐based
frameworks
during
AMPLab
period
(Spark)
▪ major
focus
on
reinforcement
learning
Ray:
a
distributed
execution
framework
for
emerging
AI
applications
25. Increasing
role
of
the
hardware
interface
25
▪ earlier
generations
of
virtualization
abstracted
away
hardware;
however,
containers
allow
direct
access
▪ with
DL,
application
software
must
access
the
latest
hardware
features
directly
–
to
be
competitive
▪ vendors
anticipate
adv.
math
needs
for
low-‐level
hardware,
looking
beyond
DL
–
e.g.,
multi-‐linear
algebra
libraries
▪ Scaling
machine
learning
(O’Reilly
Data
Show,
21:43)
Reza
Zadeh
Stanford
/
Matroid
26. Emerging
themes:
transfer
learning
▪ transfer
learning:
when
you
can
solve
a
task
well,
transfer
understanding
to
solve
related
problems
▪ remove
final
classification
layer,
then
extract
next-‐to-‐last
layer
of
a
CNN:
tensorflow.org/tutorials/image_recognition
▪ leverage
a
network
pre-‐trained
on
a
large
dataset:
blog.keras.io/building-‐powerful-‐image-‐classification-‐
models-‐using-‐very-‐little-‐data.html
26
27. Emerging
themes:
GANs
▪ generative
adversarial
networks:
neural
networks
compete
against
each
other
in
a
zero-‐sum
game
▪ example:
CycleGAN
(see
AI
NY
2017)
27
29. LSTM
used
to
generate
content
29
Long
short-‐term
memory
(LSTM)
allows
recurrent
neural
networks
to
learn
sequences
of
data,
such
as
in
streams
of
voice
or
text.
Imagine
feeding
scripts
(semi-‐structured
data)
from
a
film
genre
through
an
LSTM,
then
generating
new
output…
30. LSTM
used
to
generate
content
30
http://benjamin.wtf/
Sunspring
It’s
No
Game
31. LSTM
in
music
composition
/
performance
31
https://github.com/IraKorshunova/folk-‐rnn
35. Peer
Teaching
through
a
range
of
Media
▪ books,
videos
▪ live
online
courses
▪ conferences
▪ AMAs
▪ computable
content
▪ case
studies
▪ articles
▪ podcast
interviews
▪ chat
forums
35
39. Key
insight
for
AI
in
Media:
▪ any
content
which
can
represented
as
text
can
be
parsed
by
NLP,
then
manipulated
by
available
AI
tooling
▪ labeled
images
get
really
interesting
▪ text
or
images
within
a
context
have
inherent
structure
▪ representation
of
that
kind
of
structure
is
rare
in
the
Media
vertical
–
so
far
39
41. Ontology
▪ provides
context
which
Deep
Learning
lacks
▪ aka,
“knowledge
graph”
–
a
computable
thesaurus
▪ maps
the
semantics
of
business
relationships
▪ S/V/O:
“nouns”,
some
“verbs”,
a
few
“adjectives”
▪ conversational
interfaces
(e.g.,
Google
Assistant)
improve
UX
by
importing
ontologies
▪ the
hard
part,
a
relatively
expensive
investment
41
42. Which
parts
do
people
or
machines
do
best?
42
team
goal:
maintain
structural
correspondence
between
the
layers
big
win
for
AI:
inferences
across
the
graph
human
scale
primary
structure
control
points
testability
machine
generated
data
products
~80%
of
the
graph
46. PyTextRank
46
TextRank
(R
Mihalcea,
P
Tarau,
2004)
a
graph
algorithm
that
extracts
key
phrases
and
summarizes
texts
–
for
NLP
which
is
improved
over
use
of
keywords,
n-‐grams,
etc.
▪ construct
a
graph
from
a
paragraph
of
text
▪ run
PageRank
on
that
graph
▪ extract
the
highly
ranked
phrases
Python
implementation
atop
spaCy,
NetworkX,
datasketch:
▪ https://pypi.python.org/pypi/pytextrank/
48. Working
with
text
and
NLP
48
▪ parsing
▪ named
entity
recognition
▪ vector
embedding
▪ smarter
indexing
▪ summarization
(especially
video)
▪ semantic
similarity
to
suggest
curriculum
▪ speed
development
of
assessments
▪ query
expansion
▪ amending
ontology
49. A
plug
for
InnerSource…
49
We
thought
the
introduction
of
data
science
had
run
headlong
into
enterprise
silos
and
lingering
tech
debt.
As
if!!
Introduction
of
AI
exacerbates
that
problem
even
more.
Suggested
responses:
▪ InnerSourceCommons.org
open
source
practices
within
enterprise
▪ design
patterns
for
working
across
silos
▪ think:
“good
house
rules
for
guests”
as
other
teams
submit
PRs
on
your
code
repos
51. A
generational
shift?
▪ We’re
12
years
beyond
the
introduction
of
YouTube
…
anyone
raising
tweens
now
probably
knows
about
YouTubers
▪ Below
a
certain
age
demographic,
people
tend
to
rely
more
on
video
and
audio
sources
for
information,
while
perhaps
print
is
gaining
more
for
entertainment.
Mobile
certainly
has
huge
impact
there.
51
56. Active
learning
▪ special
case
of
semi-‐supervised
machine
learning
▪ send
difficult
calls
/
edge
cases
to
experts;
let
algorithms
handle
routine
decisions
▪ works
well
in
use
cases
which
have
lots
of
inexpensive,
unlabeled
data
▪ e.g.,
abundance
of
content
to
be
classified,
where
the
cost
of
labeling
is
the
expense
▪ https://en.wikipedia.org/wiki/
Active_learning_(machine_learning)
56
57.
58. Active
learning
Data
preparation
in
the
age
of
deep
learning
oreilly.com/ideas/data-‐preparation-‐in-‐the-‐
age-‐of-‐deep-‐learning
Luke
Biewald
CrowdFlower
O’Reilly
Data
Show,
2017-‐05-‐04
send
human
workers
cases
where
machine
learning
algorithms
signal
uncertainty
(low
probability
scores)
or
when
your
ensemble
of
machine
learning
algorithms
signals
disagreement
58
59. Human-‐in-‐the-‐loop
design
pattern
Building
a
business
that
combines
human
experts
and
data
science
oreilly.com/ideas/building-‐a-‐business-‐that-‐
combines-‐human-‐experts-‐and-‐data-‐science-‐2
Eric
Colson
StitchFix
O’Reilly
Data
Show,
2016-‐01-‐28
“what
machines
can’t
do
are
things
around
cognition,
things
that
have
to
do
with
ambient
information,
or
appreciation
of
aesthetics,
or
even
the
ability
to
relate
to
another
human”
59
60. Weak
supervision
Creating
large
training
data
sets
quickly
oreilly.com/ideas/creating-‐large-‐
training-‐data-‐sets-‐quickly
Alex
Ratner
Stanford
O’Reilly
Data
Show,
2017-‐06-‐08
Snorkel:
“data
programming”
as
another
instance
of
human-‐in-‐the-‐loop
github.com/HazyResearch/snorkel
conferences.oreilly.com/strata/strata-‐ny/
public/schedule/detail/61849
60
61. Collaboration
through
Jupyter
61
Notebooks
get
used
to
manage
ML
pipelines,
where
machines
+
people
collaborate
on
docs
▪ “Human-‐in-‐the-‐loop
design
pattern”
talk
@
JupyterCon
NY
2017
▪ experts
adjust
parameters
in
ML
pipelines
▪ machines
write
structured
“logs”
of
ML
modeling
and
evaluation
▪ experts
run
`jupyter
notebook`
via
SSH
tunnel
for
remote
monitoring
and
updates
▪ https://pypi.python.org/pypi/nbtransom
63. Collaboration
through
Jupyter
▪ running
notebooks
via
SSH
tunnel
removes
the
need
for
dedicated
UIs
▪ this
work
anticipates
upcoming
collaborative
document
features
in
JupyterLab:
Realtime
collaboration
for
JupyterLab
using
Google
Drive
Ian
Rose
UC
Berkeley
64. Expert
review
▪ ML
pipelines
report
results:
recognizing
content,
adding
annotations,
requesting
more
examples
when
“confused”
▪ Human-‐in-‐the-‐loop
experts
–
potentially,
Customer
Service
–
review
decisions,
especially
edge
cases,
then
train
through
examples
▪ The
system
iterates
64
65. What’s
the
point
of
using
AI
in
Media?
▪ more
work,
quicker,
than
could
be
performed
by
editors
–
who
are
already
super-‐busy
people
▪ exceeding
human
parity,
as
a
benchmark
▪ helps
relieve
pressure
on
organizations,
as
learning
curves
accelerate
▪ augments
some
of
our
most
valuable
experts,
so
they
can
get
more
done
65
66. Human-‐in-‐the-‐loop
as
a
management
strategy
66
personal
op-‐ed:
the
“game”
isn’t
to
replace
people
–
instead
it’s
about
leveraging
AI
to
augment
staff,
so
organizations
can
retain
people
with
valuable
domain
expertise,
making
their
contributions
and
expertise
even
more
vital
70. 70
Learn
Alongside
Innovators
Just
Enough
Math Building
Data
Science
Teams
Hylbert-‐Speys How
Do
You
Learn?
updates,
reviews,
conference
summaries…
liber118.com/pxn/
@pacoid