<s>
Automated	O
Artificial	B-Application
Intelligence	I-Application
(	O
AutoAI	B-Algorithm
)	O
is	O
a	O
variation	O
of	O
the	O
automated	B-General_Concept
machine	I-General_Concept
learning	I-General_Concept
,	O
or	O
AutoML	B-General_Concept
,	O
technology	O
,	O
which	O
extends	O
the	O
automation	O
of	O
model	O
building	O
towards	O
automation	O
of	O
the	O
full	O
life	O
cycle	O
of	O
a	O
machine	O
learning	O
model	O
.	O
</s>
<s>
It	O
applies	O
intelligent	B-Application
automation	I-Application
to	O
the	O
task	O
of	O
building	O
predictive	B-General_Concept
machine	O
learning	O
models	O
by	O
preparing	O
data	O
for	O
training	O
,	O
identifying	O
the	O
best	O
type	O
of	O
model	O
for	O
the	O
given	O
data	O
,	O
then	O
choosing	O
the	O
features	O
,	O
or	O
columns	O
of	O
data	O
,	O
that	O
best	O
support	O
the	O
problem	O
the	O
model	O
is	O
solving	O
.	O
</s>
<s>
Automated	O
artificial	B-Application
intelligence	I-Application
can	O
also	O
be	O
applied	O
to	O
making	O
sure	O
the	O
model	O
does	O
not	O
have	O
inherent	O
bias	O
and	O
automating	O
the	O
tasks	O
for	O
continuous	O
improvement	O
of	O
the	O
model	O
.	O
</s>
<s>
Managing	O
an	O
AutoAI	B-Algorithm
model	O
requires	O
frequent	O
monitoring	O
and	O
updating	O
,	O
managed	O
by	O
a	O
process	O
known	O
as	O
model	O
operations	O
,	O
or	O
ModelOps	O
.	O
</s>
<s>
The	O
Automated	B-General_Concept
Machine	I-General_Concept
Learning	I-General_Concept
and	O
Data	O
Science	O
Team	O
(	O
AMLDS	O
)	O
,	O
a	O
small	O
team	O
within	O
IBM	O
Research	O
,	O
which	O
was	O
formed	O
to	O
“	O
apply	O
techniques	O
from	O
artificial	B-Application
intelligence	I-Application
(	O
AI	B-Application
)	O
,	O
machine	O
learning	O
(	O
ML	O
)	O
,	O
and	O
data	B-General_Concept
management	I-General_Concept
to	O
accelerate	O
and	O
optimize	O
the	O
creation	O
of	O
machine	O
learning	O
and	O
data	O
science	O
workflows	O
,	O
”	O
is	O
credited	O
with	O
advancing	O
the	O
development	O
of	O
AutoAI	B-Algorithm
.	O
</s>
<s>
A	O
typical	O
use	O
case	O
for	O
AutoAI	B-Algorithm
would	O
be	O
training	O
a	O
model	O
to	O
predict	O
how	O
customers	O
might	O
respond	O
to	O
a	O
sales	O
incentive	O
.	O
</s>
<s>
Prior	O
to	O
AutoML	B-General_Concept
,	O
data	O
scientists	O
had	O
to	O
build	O
these	O
predictive	B-General_Concept
models	I-General_Concept
by	O
hand	O
,	O
testing	O
various	O
combinations	O
of	O
algorithms	O
,	O
then	O
testing	O
to	O
see	O
how	O
predictions	O
compared	O
to	O
actual	O
results	O
.	O
</s>
<s>
Where	O
AutoML	B-General_Concept
automated	O
some	O
of	O
the	O
process	O
of	O
preparing	O
the	O
data	O
for	O
training	O
,	O
applying	O
algorithms	O
to	O
process	O
the	O
data	O
and	O
then	O
further	O
optimizing	O
the	O
results	O
,	O
AutoAI	B-Algorithm
provides	O
greater	O
intelligent	B-Application
automation	I-Application
that	O
allows	O
for	O
testing	O
significantly	O
more	O
combinations	O
of	O
factors	O
to	O
generate	O
model	O
candidate	O
pipelines	O
that	O
more	O
accurately	O
reflect	O
and	O
address	O
the	O
problem	O
being	O
solved	O
.	O
</s>
<s>
In	O
the	O
data	O
pre-processing	B-General_Concept
stage	O
,	O
AutoAI	B-Algorithm
applies	O
numerous	O
algorithms	O
,	O
or	O
estimators	O
,	O
to	O
analyze	O
,	O
clean	O
(	O
for	O
example	O
,	O
remove	O
redundant	O
information	O
or	O
impute	O
missing	O
data	O
)	O
,	O
and	O
prepare	O
structured	O
raw	O
data	O
for	O
machine	O
learning	O
(	O
ML	O
)	O
.	O
</s>
<s>
For	O
example	O
,	O
if	O
there	O
are	O
only	O
two	O
types	O
of	O
data	O
in	O
a	O
prediction	O
column	O
,	O
AutoAI	B-Algorithm
prepares	O
to	O
build	O
a	O
binary	O
classification	O
model	O
.	O
</s>
<s>
If	O
there	O
is	O
an	O
unknowable	O
set	O
of	O
possible	O
answers	O
,	O
AutoAI	B-Algorithm
prepares	O
a	O
regression	O
model	O
,	O
which	O
employs	O
a	O
different	O
set	O
of	O
algorithms	O
,	O
or	O
problem-solving	O
transformations	O
.	O
</s>
<s>
AutoAI	B-Algorithm
ranks	O
after	O
testing	O
candidate	O
algorithms	O
against	O
small	O
sub-sets	O
of	O
the	O
information	O
,	O
increasing	O
the	O
size	O
of	O
the	O
subset	O
gradually	O
for	O
the	O
algorithms	O
that	O
turns	O
most	O
promising	O
to	O
reach	O
at	O
the	O
best	O
match	O
.	O
</s>
<s>
This	O
process	O
of	O
iterative	O
and	O
incremental	O
machine	O
learning	O
is	O
what	O
sets	O
AutoAI	B-Algorithm
apart	O
from	O
earlier	O
versions	O
of	O
AutoML	B-General_Concept
.	O
</s>
<s>
Feature	B-General_Concept
engineering	I-General_Concept
transforms	O
the	O
raw	O
data	O
into	O
the	O
combination	O
that	O
represents	O
the	O
problem	O
to	O
arrive	O
at	O
the	O
best	O
accurate	O
prediction	O
.	O
</s>
<s>
AutoAI	B-Algorithm
automates	O
the	O
consideration	O
of	O
numerous	O
feature	O
construction	O
options	O
in	O
a	O
non-exhaustive	O
,	O
structured	O
manner	O
,	O
meanwhile	O
progressively	O
maximizing	O
the	O
accuracy	O
of	O
model	O
using	O
reinforcement	O
learning	O
.	O
</s>
<s>
Finally	O
,	O
AutoAI	B-Algorithm
applies	O
the	O
hyperparameter	B-General_Concept
optimization	I-General_Concept
step	O
to	O
refine	O
and	O
advance	O
the	O
best	O
performing	O
model	O
pipelines	O
.	O
</s>
<s>
In	O
August	O
2017	O
,	O
AMLDS	O
announced	O
that	O
they	O
were	O
researching	O
the	O
use	O
of	O
automated	B-General_Concept
feature	I-General_Concept
engineering	I-General_Concept
to	O
eliminate	O
guesswork	O
in	O
data	O
science	O
.	O
</s>
<s>
Called	O
“	O
Learning-based	O
Feature	B-General_Concept
Engineering	I-General_Concept
,	O
”	O
their	O
method	O
learned	O
the	O
correlations	O
between	O
feature	O
distributions	O
,	O
target	O
distributions	O
,	O
and	O
transformations	O
,	O
built	O
meta-models	O
that	O
used	O
past	O
observations	O
to	O
predict	O
viable	O
transformations	O
,	O
and	O
generalized	O
thousands	O
of	O
data	O
sets	O
spanning	O
different	O
domains	O
.	O
</s>
<s>
In	O
2018	O
,	O
IBM	O
Research	O
announced	O
Deep	B-Algorithm
Learning	I-Algorithm
as	O
a	O
Service	O
,	O
which	O
opened	O
popular	O
deep	B-Algorithm
learning	I-Algorithm
libraries	O
such	O
as	O
Caffe	O
,	O
Torch	O
and	O
TensorFlow	B-Language
,	O
to	O
developers	O
in	O
the	O
cloud	O
.	O
</s>
<s>
He	O
found	O
out	O
and	O
decided	O
to	O
be	O
ready	O
for	O
IBM	O
AI	B-Application
and	O
data	O
science	O
platforms	O
like	O
IBM	B-Application
Watson	I-Application
.	O
</s>
<s>
In	O
December	O
2018	O
,	O
IBM	O
Research	O
announced	O
NeuNetS	O
,	O
a	O
new	O
capability	O
that	O
automated	O
neural	O
network	O
model	O
synthesis	O
as	O
part	O
of	O
automated	O
AI	B-Application
model	O
development	O
and	O
deployment	O
.	O
</s>
<s>
proposed	O
a	O
method	O
for	O
AutoML	B-General_Concept
that	O
used	O
the	O
alternating	O
direction	O
method	O
of	O
multipliers	O
(	O
ADMM	O
)	O
to	O
configure	O
multiple	O
stages	O
of	O
an	O
ML	O
pipeline	O
,	O
such	O
as	O
transformations	O
,	O
feature	B-General_Concept
engineering	I-General_Concept
and	O
selection	O
,	O
and	O
predictive	B-General_Concept
modeling	I-General_Concept
.	O
</s>
<s>
2019	O
was	O
the	O
year	O
that	O
AutoML	B-General_Concept
became	O
more	O
widely	O
discussed	O
as	O
a	O
concept	O
.	O
</s>
<s>
“	O
The	O
Forrester	O
New	O
Wave™	O
:	O
Automation-Focused	O
Machine	O
Learning	O
Solutions	O
,	O
Q2	O
2019	O
,	O
”	O
evaluated	O
AutoML	B-General_Concept
solutions	O
and	O
found	O
that	O
the	O
more	O
powerful	O
versions	O
offered	O
feature	B-General_Concept
engineering	I-General_Concept
.	O
</s>
<s>
A	O
Gartner	O
Technical	O
Professional	O
Advice	O
report	O
from	O
August	O
2019	O
reported	O
that	O
,	O
based	O
on	O
their	O
research	O
,	O
AutoML	B-General_Concept
could	O
augment	O
data	O
science	O
and	O
machine	O
learning	O
.	O
</s>
<s>
They	O
described	O
AutoML	B-General_Concept
as	O
the	O
automation	O
of	O
data	O
preparation	O
,	O
feature	B-General_Concept
engineering	I-General_Concept
and	O
model	O
engineering	O
tasks	O
.	O
</s>
<s>
AutoAI	B-Algorithm
is	O
the	O
evolution	O
of	O
AutoML	B-General_Concept
.	O
</s>
<s>
One	O
of	O
AutoAI	B-Algorithm
's	O
principal	O
inventors	O
,	O
Jean-Francois	O
Puget	O
,	O
PhD	O
,	O
describes	O
it	O
as	O
automatically	O
performing	O
data	O
preparation	O
,	O
feature	B-General_Concept
engineering	I-General_Concept
,	O
machine	O
learning	O
algorithm	O
selection	O
,	O
and	O
hyper-parameter	B-General_Concept
optimization	I-General_Concept
to	O
find	O
the	O
best	O
possible	O
machine	O
learning	O
model	O
.	O
</s>
<s>
The	O
hyper-parameter	B-General_Concept
optimization	I-General_Concept
algorithm	O
used	O
in	O
AutoAI	B-Algorithm
differs	O
from	O
the	O
hyper-parameter	O
tuning	O
of	O
AutoML	B-General_Concept
.	O
</s>
<s>
Research	O
scientists	O
at	O
IBM	O
Research	O
published	O
a	O
paper	O
"	O
Towards	O
Automating	O
the	O
AI	B-Application
Operations	O
Lifecycle	O
"	O
,	O
which	O
describes	O
the	O
advantages	O
and	O
available	O
technologies	O
for	O
automating	O
more	O
of	O
the	O
process	O
,	O
with	O
the	O
goal	O
of	O
limiting	O
the	O
human	O
involvement	O
required	O
to	O
build	O
,	O
test	O
,	O
and	O
maintain	O
a	O
machine	O
learning	O
application	O
.	O
</s>
<s>
Rather	O
,	O
a	O
more	O
transparent	O
and	O
interpretable	O
AutoAI	B-Algorithm
design	O
is	O
the	O
key	O
to	O
gain	O
trust	O
from	O
human	O
users	O
,	O
but	O
such	O
design	O
itself	O
is	O
quite	O
a	O
challenge	O
.	O
</s>
<s>
Winner	O
,	O
Best	O
Innovation	O
in	O
Intelligent	B-Application
Automation	I-Application
Award	O
at	O
the	O
AIconics	O
AI	B-Application
Summit	O
(	O
2019	O
)	O
,	O
San	O
Francisco	O
.	O
</s>
