<s>
Flexibility	O
is	O
important	O
because	O
each	O
learning	O
algorithm	O
is	O
based	O
on	O
a	O
set	O
of	O
assumptions	O
about	O
the	O
data	O
,	O
its	O
inductive	B-General_Concept
bias	I-General_Concept
.	O
</s>
<s>
This	O
poses	O
strong	O
restrictions	O
on	O
the	O
use	O
of	O
machine	O
learning	O
or	O
data	B-Application
mining	I-Application
techniques	O
,	O
since	O
the	O
relationship	O
between	O
the	O
learning	O
problem	O
(	O
often	O
some	O
kind	O
of	O
database	O
)	O
and	O
the	O
effectiveness	O
of	O
different	O
learning	O
algorithms	O
is	O
not	O
yet	O
understood	O
.	O
</s>
<s>
Critiques	O
of	O
meta	B-General_Concept
learning	I-General_Concept
approaches	O
bear	O
a	O
strong	O
resemblance	O
to	O
the	O
critique	O
of	O
metaheuristic	B-Algorithm
,	O
a	O
possibly	O
related	O
problem	O
.	O
</s>
<s>
A	O
good	O
analogy	O
to	O
meta-learning	B-General_Concept
,	O
and	O
the	O
inspiration	O
for	O
Jürgen	O
Schmidhuber	O
's	O
early	O
work	O
(	O
1987	O
)	O
and	O
Yoshua	O
Bengio	O
et	O
al	O
.	O
</s>
<s>
In	O
an	O
open-ended	O
hierarchical	O
meta	B-General_Concept
learning	I-General_Concept
system	O
using	O
genetic	B-Algorithm
programming	I-Algorithm
,	O
better	O
evolutionary	O
methods	O
can	O
be	O
learned	O
by	O
meta	O
evolution	O
,	O
which	O
itself	O
can	O
be	O
improved	O
by	O
meta	O
meta	O
evolution	O
,	O
etc	O
.	O
</s>
<s>
See	O
also	O
Ensemble	B-Algorithm
learning	I-Algorithm
.	O
</s>
<s>
A	O
proposed	O
definition	O
for	O
a	O
meta	B-General_Concept
learning	I-General_Concept
system	O
combines	O
three	O
requirements	O
:	O
</s>
<s>
Learning	B-General_Concept
bias	I-General_Concept
must	O
be	O
chosen	O
dynamically	O
.	O
</s>
<s>
Bias	O
refers	O
to	O
the	O
assumptions	O
that	O
influence	O
the	O
choice	O
of	O
explanatory	O
hypotheses	O
and	O
not	O
the	O
notion	O
of	O
bias	O
represented	O
in	O
the	O
bias-variance	B-General_Concept
dilemma	I-General_Concept
.	O
</s>
<s>
Meta	B-General_Concept
learning	I-General_Concept
is	O
concerned	O
with	O
two	O
aspects	O
of	O
learning	B-General_Concept
bias	I-General_Concept
.	O
</s>
<s>
Model-based	O
meta-learning	B-General_Concept
models	O
updates	O
its	O
parameters	O
rapidly	O
with	O
a	O
few	O
training	O
steps	O
,	O
which	O
can	O
be	O
achieved	O
by	O
its	O
internal	O
architecture	O
or	O
controlled	O
by	O
another	O
meta-learner	O
model	O
.	O
</s>
<s>
A	O
Memory-Augmented	O
Neural	B-Architecture
Network	I-Architecture
,	O
or	O
MANN	O
for	O
short	O
,	O
is	O
claimed	O
to	O
be	O
able	O
to	O
encode	O
new	O
information	O
quickly	O
and	O
thus	O
to	O
adapt	O
to	O
new	O
tasks	O
after	O
only	O
a	O
few	O
examples	O
.	O
</s>
<s>
The	O
core	O
idea	O
in	O
metric-based	O
meta-learning	B-General_Concept
is	O
similar	O
to	O
nearest	B-General_Concept
neighbors	I-General_Concept
algorithms	O
,	O
which	O
weight	O
is	O
generated	O
by	O
a	O
kernel	O
function	O
.	O
</s>
<s>
Siamese	B-Algorithm
neural	I-Algorithm
network	I-Algorithm
is	O
composed	O
of	O
two	O
twin	O
networks	O
whose	O
output	O
is	O
jointly	O
trained	O
.	O
</s>
<s>
During	O
meta-learning	B-General_Concept
,	O
it	O
learns	O
to	O
learn	O
a	O
deep	O
distance	O
metric	O
to	O
compare	O
a	O
small	O
number	O
of	O
images	O
within	O
episodes	O
,	O
each	O
of	O
which	O
is	O
designed	O
to	O
simulate	O
the	O
few-shot	O
setting	O
.	O
</s>
<s>
Compared	O
to	O
recent	O
approaches	O
for	O
few-shot	O
learning	O
,	O
they	O
reflect	O
a	O
simpler	O
inductive	B-General_Concept
bias	I-General_Concept
that	O
is	O
beneficial	O
in	O
this	O
limited-data	O
regime	O
,	O
and	O
achieve	O
satisfied	O
results	O
.	O
</s>
<s>
What	O
optimization-based	O
meta-learning	B-General_Concept
algorithms	O
intend	O
for	O
is	O
to	O
adjust	O
the	O
optimization	O
algorithm	O
so	O
that	O
the	O
model	O
can	O
be	O
good	O
at	O
learning	O
with	O
a	O
few	O
examples	O
.	O
</s>
<s>
LSTM-based	O
meta-learner	O
is	O
to	O
learn	O
the	O
exact	O
optimization	O
algorithm	O
used	O
to	O
train	O
another	O
learner	O
neural	B-Architecture
network	I-Architecture
classifier	B-General_Concept
in	O
the	O
few-shot	O
regime	O
.	O
</s>
<s>
The	O
parametrization	O
allows	O
it	O
to	O
learn	O
appropriate	O
parameter	O
updates	O
specifically	O
for	O
the	O
scenario	O
where	O
a	O
set	O
amount	O
of	O
updates	O
will	O
be	O
made	O
,	O
while	O
also	O
learning	O
a	O
general	O
initialization	O
of	O
the	O
learner	O
(	O
classifier	B-General_Concept
)	O
network	O
that	O
allows	O
for	O
quick	O
convergence	O
of	O
training	O
.	O
</s>
<s>
MAML	O
,	O
short	O
for	O
Model-Agnostic	O
Meta-Learning	B-General_Concept
,	O
is	O
a	O
fairly	O
general	O
optimization	O
algorithm	O
,	O
compatible	O
with	O
any	O
model	O
that	O
learns	O
through	O
gradient	O
descent	O
.	O
</s>
<s>
Reptile	O
is	O
a	O
remarkably	O
simple	O
meta-learning	B-General_Concept
optimization	O
algorithm	O
,	O
given	O
that	O
both	O
of	O
its	O
components	O
rely	O
on	O
meta-optimization	B-Algorithm
through	O
gradient	O
descent	O
and	O
both	O
are	O
model-agnostic	O
.	O
</s>
<s>
Some	O
approaches	O
which	O
have	O
been	O
viewed	O
as	O
instances	O
of	O
meta	B-General_Concept
learning	I-General_Concept
:	O
</s>
<s>
Recurrent	B-Algorithm
neural	I-Algorithm
networks	I-Algorithm
(	O
RNNs	O
)	O
are	O
universal	O
computers	O
.	O
</s>
<s>
In	O
1993	O
,	O
Jürgen	O
Schmidhuber	O
showed	O
how	O
"	O
self-referential	O
"	O
RNNs	O
can	O
in	O
principle	O
learn	O
by	O
backpropagation	B-Algorithm
to	O
run	O
their	O
own	O
weight	O
change	O
algorithm	O
,	O
which	O
may	O
be	O
quite	O
different	O
from	O
backpropagation	B-Algorithm
.	O
</s>
<s>
Conwell	O
built	O
a	O
successful	O
supervised	O
meta	O
learner	O
based	O
on	O
Long	B-Algorithm
short-term	I-Algorithm
memory	I-Algorithm
RNNs	O
.	O
</s>
<s>
It	O
learned	O
through	O
backpropagation	B-Algorithm
a	O
learning	O
algorithm	O
for	O
quadratic	O
functions	O
that	O
is	O
much	O
faster	O
than	O
backpropagation	B-Algorithm
.	O
</s>
<s>
Researchers	O
at	O
Deepmind	B-Application
(	O
Marcin	O
Andrychowicz	O
et	O
al	O
.	O
)	O
</s>
<s>
An	O
extreme	O
type	O
of	O
Meta	O
Reinforcement	O
Learning	O
is	O
embodied	O
by	O
the	O
Gödel	B-General_Concept
machine	I-General_Concept
,	O
a	O
theoretical	O
construct	O
which	O
can	O
inspect	O
and	O
modify	O
any	O
part	O
of	O
its	O
own	O
software	O
which	O
also	O
contains	O
a	O
general	O
theorem	B-Application
prover	I-Application
.	O
</s>
<s>
Model-Agnostic	O
Meta-Learning	B-General_Concept
(	O
MAML	O
)	O
was	O
introduced	O
in	O
2017	O
by	O
Chelsea	O
Finn	O
et	O
al	O
.	O
</s>
<s>
Boosting	B-Algorithm
is	O
related	O
to	O
stacked	O
generalisation	O
,	O
but	O
uses	O
the	O
same	O
algorithm	O
multiple	O
times	O
,	O
where	O
the	O
examples	O
in	O
the	O
training	O
data	O
get	O
different	O
weights	O
over	O
each	O
run	O
.	O
</s>
<s>
Dynamic	O
bias	O
selection	O
works	O
by	O
altering	O
the	O
inductive	B-General_Concept
bias	I-General_Concept
of	O
a	O
learning	O
algorithm	O
to	O
match	O
the	O
given	O
problem	O
.	O
</s>
<s>
Inductive	B-General_Concept
transfer	I-General_Concept
studies	O
how	O
the	O
learning	O
process	O
can	O
be	O
improved	O
over	O
time	O
.	O
</s>
<s>
Other	O
approaches	O
using	O
metadata	O
to	O
improve	O
automatic	O
learning	O
are	O
learning	B-Algorithm
classifier	I-Algorithm
systems	I-Algorithm
,	O
case-based	O
reasoning	O
and	O
constraint	B-Application
satisfaction	I-Application
.	O
</s>
<s>
Some	O
initial	O
,	O
theoretical	O
work	O
has	O
been	O
initiated	O
to	O
use	O
Applied	O
Behavioral	O
Analysis	O
as	O
a	O
foundation	O
for	O
agent-mediated	O
meta-learning	B-General_Concept
about	O
the	O
performances	O
of	O
human	O
learners	O
,	O
and	O
adjust	O
the	O
instructional	O
course	O
of	O
an	O
artificial	O
agent	O
.	O
</s>
<s>
AutoML	B-General_Concept
such	O
as	O
Google	O
Brain	O
's	O
"	O
AI	O
building	O
AI	O
"	O
project	O
,	O
which	O
according	O
to	O
Google	O
briefly	O
exceeded	O
existing	O
ImageNet	B-General_Concept
benchmarks	O
in	O
2017	O
.	O
</s>
