►
Description
Lucas Souza discusses few-shot learning and relevant benchmarks used to measure performance in these settings. This is part of a series of research meetings aimed at reviewing common training paradigms and benchmarks in machine learning, and their relation to the Thousand Brains Theory of Intelligence.
- - - - -
Numenta is leading the new era of machine intelligence. Our deep experience in theoretical neuroscience research has led to tremendous discoveries on how the brain works. We have developed a cohesive theory, core software technology, and numerous software applications all based on principles of the neocortex. Our innovative work delivers breakthrough capabilities and demonstrates that a computing approach based on biological learning principles can do things that today’s programmed computers cannot do.
Subscribe to our News Digest for the latest news about neuroscience and artificial intelligence:
https://tinyurl.com/NumentaNewsDigest
Subscribe to our Newsletter for the latest Numenta updates:
https://tinyurl.com/NumentaNewsletter
Our Social Media:
https://twitter.com/Numenta
https://www.facebook.com/OfficialNumenta
https://www.linkedin.com/company/numenta
Our Open Source Resources:
https://github.com/numenta
https://discourse.numenta.org/
Our Website:
https://numenta.com/
A
But
there
are
a
lot
of
intermediate
steps
which
are
are
not
yet
clear.
How
do
we
put
that
into
a
machine
learning
code?
So
I
think,
having
intermediate
benchmarks
might
be
a
good
way
to
go
and
that's
kind
of
what
we're
doing
now
right.
We
have
a
sparse
networks
and
supervised
learning
and
now
we're
moving
to
few
shot
and
continue
learning,
and
I
figure
we'll
continue
move
through
that
path
until
we
get
to
our
end
goal.
A
A
What
would
an
ideal
benchmark
look
like,
so
I
think
what
we
want
and
then
I
think
this
slide.
Please
stop
me
anytime
and
correct
me
if
I'm
wrong,
so
it's
just
my
ideas
on
it,
so
you
want
to
learn
a
mod
of
the
word
from
a
continuous
screen
of
sensory
input
by
interacting
with
an
environmental
behavior
with
little
supervision
or
reward
signals.
A
To
hear
it
okay,
so
so
what
what
I
want
to
do
in
this
project?
So
I
want
to
break
this
down
into
different
training
paradigms
and
benchmarks
and
see
how
are
these
problems
being
addressed
in
machine
learning,
and
so
I
wanted
to
start
with
the
last
one,
with
little
supervision
or
revit
signals.
And
specifically,
I
wanted
to
talk
about
flu
shot
learning
because
that's
the
problem.
B
It
it
I'm
just
saying
it's
not
clear
that
belongs
there.
That's
an
attribute
that
applies
to
to
the
brain.
Clearly
I
guess
you
could
call
it.
I
mean,
because
it's
not
really
a
supervisor
if
it's
self-supervised,
if
we
think
the
brain
is
largely
sub-supervised,
then
it's
a
separate
attribute.
It's
fine!
It's!
I
think
it's
a
good
thing
to
focus
on
I'm
just
pointing
out
that
I
don't
necessarily
include
it
under
the
little
supervision.
It's
it's
more
of
just
a
self-supervision.
B
B
I
understand
I'm
just
saying
in
terms
of
our
world
the
four
lines
above
that
we
just
said
were
great.
I
don't
think
anything
said
that
that
was
you
didn't
bring
out
the
fact
that
it's
it
requires
a
few.
You
know
that
it's
a
very
rapid
learning
system,
so
it's
all
right.
It's
great!
It's
a
separate
little
element.
B
A
A
And
the
idea
is
that
you're
gonna
leverage,
the
information
you
get
from
these
other
classes,
you're
going
to
learn
how
to
learn
in
a
few
steps.
And
you
need
to
leverage
this.
The
collective
of
tests,
not
just
one.
So
if
it
was
just
one
class
and
just
a
few
samples,
would
be
a
completely
different
problem
than
having
50
different
classes
and
like
20
samples
per
class.
So.
A
A
Well,
it's
not
the
same.
That
would
be
more
more
towards
well,
it's
kind
of
that
yeah.
What
I
mean.
It's
not
just
learning
one
problem
and
you
only
have
those
20
samples
and
that's
it
right.
You
have
access
to
several
tasks
which
belong
to
the
same
distribution,
and
you
can
leverage
that
in
order
to
learn
in
that
particular
task
that
that's
that's
kind
of
where
I'm
going
so.
A
C
So
it's
sort
of
similar
to
the
intuition
I
mean
in
biology.
I
mean
we
sort
of
grow
up
learning
or
just
about
the
world,
all
sorts
of
stuff
about
the
world,
and
then
we
can
learn
a
new
thing
very
quickly
by
leveraging
what
we've
learned
in
the
past
is
that:
are
they
trying
to
get
at
something
like
this,
that,
by
learning
by
learning
a
lot
of
different
classes,
it's
like
learning
stuff
in
the
background
that
might
be
help
you
in
learning
something
new.
D
B
If
I
were
to
translate
this
into
language
of
the
thousand
brains
theory,
I
would
say
maybe
something
like
the
following.
I
would
say
a
column
has
to
discover
the
right
reference
frames,
motor
motor
and
the
motor
behaviors
and
the
representation
of
space
in
the
reference
frame
and
once
a
column
has
learned
that
it
can
learn
new
new
structures
within
that
reference
frame
and
movements
that
can
learn
very
rapidly,
but
they're
all
going
to
be
of
the
same.
A
column
is
going
to
be
restricted
to
all
things
that
can
fit
into
that
reference
frame.
B
Where,
if
I
had
to
learn
something
completely
different,
it
required
a
different
reference
name,
then
the
column
would
be
very
slow.
It
wouldn't
do
very
well
so
that,
and
some
I'm
just
trying
to
transit
a
little
bit
more
detail
the
way
I
would
think
about
it,
not
the
way
machine
learning
people
think
about
it.
The
way
I
would
think
about
it
would
be
that
you've
got
this
once
you've
learned
a
set
of
objects
or
set
of
structures
with
within
a
single
column
or
many
columns.
D
B
E
Lucas,
when
I
look
at
that,
I
I
think
it
just
as
stated
it
reduces
down
to
learning
what's
common
within
the
class
and
what
differentiates
between
the
classes
you
know
so
that
could
devolve
down
to
just
not
not
to
a
general
learning
principle,
but
just
to
kind
of
solve
that
problem.
So
I'm
wondering
what
you
would
have
to
add
to
that
to
get
to
the
things
like
you
and
and
jeff
were
talking
about
where
there's
the
potential
for
generalization
from
there.
A
A
So
so
the
way
I'm
gonna
the
way
it's
usually
defined
machine
learning,
it's
in
the
context
of
supervised
learning.
You
have
labels,
you
don't
have
a
lot,
but
you
have
a
few
labels,
but
it
could
be
easily
extended
to
regression
problems
or
to
reinforcement,
learning
problems
where
you
just
have
like
a
few
episodes
or
a
few
trajectories,
and
I'm
just
going
to
go
over
very
quick
over
like
a
few
of
the
main
solutions
to
the
problem-
and
this
I
got
from
this.
A
This
is
a
nice
review
paper,
but
I'm
not
going
to
do
a
review
of
the
review
paper
here.
So
if
the
authors
see
this
video,
just
forgive
me
it's
not!
It's
not
going
to
be
comprehensive.
I'm
just
speaking
like
a
few
things
which
I
think
are
interesting,
so
they
propose
this
taxonomy
and
they
make
the
point
that,
in
order
to
learn
from
just
a
few
data,
you
require
some
some
prior
knowledge
which
can
be
built
either
into
the
data.
A
So
you
can
use
this
prior
knowledge
to
augment
the
supervised
experience.
So
you
make
the
data
richer.
It
can
be
built
into
the
model.
So
you
reduce
the
size
of
the
hypothesis
space
and
that's,
and
we
can
see
that,
for
example,
when
you,
when,
when
you
restrict
the
number
of
parameters
you
update
in
the
model,
so
you
make
the
model
a
lot
more.
A
A
So
building
knowledge
prior
knowledge
into
the
data
can,
for
example,
transforming
samples
from
the
training
set
and
one
example
we
have
is
auto
augment
that
can
be
seen
as
a
possible
solution
and
in
our
command.
You
learn:
use
reinforcement,
learning
to
learn
a
set
of
augmentation
policies
which
are
trained
for
a
specific
data
set,
so
the
knowledge
is
being
built
into
those
policies
that
tell
you
what
is
the
best
way
to
augmented
data.
A
Another
way
of
doing
this
is
transforming
samples
from
like
weekly
label,
like
we've,
seen
in
in
facebook
data
set
where
they
get
labels
from
instagram
and
in
instagram.
You
can
have
multiple
labels
for
the
same
image,
so
they
call
it
like
weekly
label.
It's
not
like
a
dog
or
a
cat;
it
can
be
many
things
at
once
or
just
a
non-labeled
dataset
and
one
way
of
doing
this.
I
think
we
talked
about
this
before
so
you
can
train
a
a
predictor
another
extra
person.
A
C
B
A
Yes,
yes,
so
so
these
are
different
solutions
to
the
problem
of.
I
only
have
a
few
samples.
How
do
I
solve
that?
So
one
one
of
the
solutions
is
this:
it's
not
what
we
usually
know
is
as
few
shot
learning
machine
learning
but,
like
you
said
it
doesn't
mean
you
you're
gonna,
train
you're,
gonna,
still
train
as
a
supervised
learning
model,
but
with
just
a
few
samples
you
can
extrapolate
to
a
lot
of
samples
and
then
you
can
solve
the
problem
in
this
similar
fashion.
A
But
yeah
yeah
you're
gonna
train
for
a
long
time,
yeah
or
or
you
can
also
like,
transform
samples
from
similar
large
data
sets
and
if
you're
trying
to
solve
a
problem,
you
only
have
a
few
samples
from
this
data
set.
You
can
find
similar
ones
and
then
sample
based
on
similarity
or
you
can
train
again
to
generate
new
samples
based
on
that
larger
data
set.
So
this
is
what
not
usually
what
people
know
as
few
shot
learning,
but
it's
a
it's
a
solution
to
the
problem.
A
Yeah,
it's
a
practical
solution,
and-
and
there
is
a
paper
that
came
out,
I
think,
a
month
ago
I
was
making
a
lot
of
noise.
I
forgot
the
name
and-
and
the
idea
was
around
that
same
thing-
you
know
just
it's
learning
from
fuel
shot,
but
basically
you
augment
the
data
in
such
a
way
that
it
makes
it
very
easy
to
learn
from
your
screenshot.
B
B
A
Yeah
so
a
second
way,
and
then
you
start
getting
into
the
algorithmic.
Is
you
actually
incorporate
that
knowledge
into
your
model?
And
there
are
a
few
ways
of
doing
that,
so
one
very
simple
one
is
just
parameter
sharing
like
you
share
some
layers
or
some
parameters
of
your
network
between
different
tasks,
and
I
think
we
we
see
a
little
bit
of
that
in
oml,
where
we
have
this
representation
network
and
that
representation
network
is
shared
between
tasks
or
you
can
do
parameter
time
as
well.
A
You
could
regularize
parameters
of
different
tasks
in
order
to
force
them
to
be
similar,
like
penalizing
pairwise
difference
in
same
way,
they're
using
domain
transfer.
So
these
are
ways
of
sharing
the
knowledge
between
tasks.
You
have
a
bunch
of
tests.
All
those
tests
have
just
a
small
number
of
samples,
but
you're
somehow
later
into
the
fact
that
these
tasks
are
drawn
from
the
same
distribution
in
order
to
learn
faster.
A
Okay
and
this
I
took
from
the
review
paper,
so
this
I'm
getting
to
the
most
common
models-
and
I
mean
embedding
learning-
is
one
of
the
most
common
ways
of
solving
few
shot.
Learning
where
you
try
to
learn
an
embedding
representation
in
writing
or
representation,
that's
going
to
make
it
easier
to
learn
a
few
shot
learning,
and
at
that
time
you
just.
You
can
just
measure
similarity
sorry
between
embedding
from
tests
and
from
training
samples,
and
I
get
I
don't.
A
All
right,
so
embedding
is
just
the
latent
representation,
like
you
transform
your
input
space
into
a
different
latent
space.
So
you
pass
through
a
bunch
of
transformations
like
a
neural
network
and
then
at
the
end
you
get
this
different
representation
of
the
input,
and
the
idea
is
that
this
different
representation
is
going
to
make
it
easier
for
downstream
tasks
like
field
shot
learning.
So.
A
It's
not
as
actually
pre-processing,
so
I'm
gonna
explain
one
of
these,
and
maybe
it
makes
more
clear.
So
this
is
a
matching
match
for
nora
vigneault.
I
think
it's
a
2016
paper.
It
kind
of
launched
this
new
wave
of
photoshop
learning,
and
the
idea
here
is
that
you
have
the
samples
from
from
the
training
test.
Only
you
have
a
few
samples
and
you're.
A
You
have
a
neural
network
here,
so
you're
going
to
learn
an
embedding
of
this
test
and
the
goal
here
and
an
embedding
of
the
the
sample
you're
trying
to
predict-
and
the
idea
here
is.
This-
is
a
tension
kernel
over
these
embeddings
and
you
are
constraining
the
test
in
such
a
way
that
you
want
to
learn
a
good
attention
kernel
over
this
representation
of
the
support
set.
A
So
it's
conditional
on
that
so
you're
learning,
both
at
the
same
time,
you're
learning
a
good
attention
card,
you're,
learning,
a
good
you're
learning,
a
good
representation.
That's
going
to
help
you
do
this
and
you're
learning
like
a
good
attention.
Kernel!
That's
going
to
that
will
learn
how
to
pay
attention
to
the
support
set
better
like
which
parts
of
the
support
set.
You
want
to
pay
attention
to
with
parts
of
the
representation,
and
I
don't
know
if
that
wasn't
a
good
way
of
explaining
right
well,.
A
Yes,
so
the
idea
here,
you're
conditioning
on
a
support
set
and
that
supports
that
are
those
only
you
only
have
a
few
samples
and
every
new
sample
you
get
you're
just
going
to
compare
to
all
existing
samples
you
have
and
see
which
ones
are
more
similar
to
so
the
attention
kernel.
Just
you
can
think
like
a
just
a
multiplier,
so
I'm
just
going
to
apply
a
multiplier
to
each
one
of
the
representations
and
combine
them,
and
that's
going
to
be
my
answer.
B
B
A
It
still
feels
short,
but
the
paradigm
here
is
that
you
have
many
tasks
and
you
have
many
episodes
so
you're
going
to
do
these
things.
You
only
have
like
a
few
few
shots
for
this
task,
but
you
have
many
tasks
and
you're
going
to
do
this.
Many
many
many
episodes
and
you
you
get
to
a
representation
which
is
good.
You
know
for
this
type
of
task.
You
could
say.
B
I
was
just
trying
to
understand
overall,
if
I
would
build
up
a
system
using
this
method.
Would
I
have
saved
training
time,
or
is
it
really
just
I'm
very
quickly
able
to
learn
a
new
category
like
an
inference
of
the
new
category,
but
I
still
overall,
the
total
amount
of
compute
time
to
train.
The
entire
system
is
still
large
or
not.
A
Yeah,
so
it's
still
large,
so
the
total
amount
of
time
to
train
you're,
not
training,
you're,
not
reducing,
training
time
at
all,
you're.
Just
dealing
with
the
fact
that
you
have
a
few
samples
per
task,
but
compute
time
is
too
large.
We
essentially
because
we're
going
to
do
multiple
episodes
with
multiple
tasks.
A
No,
the
task
is
the
same
thing
we
have
in
metal
learning.
The
task
is
like
a
few
classes,
for
example,
so
the
task
is
differentiate
between
these
three
classes
and
then
another
task
is
these
three
other
classes
and
and
then
it
goes
so
you
can
think
of
it
as
a
category
like
say,
a
dog
breeds
within
dogs
or
car
brands
within
time.
You
can
think
of
that
way.
A
C
A
C
The
mammal
you're,
given
you
know
a
few
examples
from
each
class
and
you're
tested
to
see
how
well
you
recognize
new
examples
from
those
classes,
and
then
you
update
the
whole
thing
so
that
the
next
time
around
the
network
is
better
at
doing
it
right.
That's
the
matter,
learning
piece
right
yeah!
This
is
that
kind
of
similar
loop
here.
A
Loop
here
so
you're
looping
within
episodes
so
that
the
outer
loop,
the
outer
epoch
we
call
better
and
what
they
call
episodes
and
the
idea
is
so
you're
shown
a
few
samples
and
you're
training.
These
two
functions
here
and
the
idea
is,
when
you
show
a
new
sample,
you're
gonna,
be
able
just
with
an
intentional
kernel
over
the
support
set.
You're
gonna
be
able
to
find
what
the
correct
class
is.
But
you're
gonna
do
this
episode
and
you
can
it's
exact
same
paradigm
as
memos
just
different
solution
to
it.
A
A
A
So
you're
moving
from
this
paradigm,
where
you
have
like
a
learning
rate
applied
to
the
gradient,
and
you
use
that
to
update
your
weights
to
this
new
paradigm,
where
you
are
actually,
you
have
a
function
and
the
step
is
going
to
be
a
function
of
the
gradients
and
the
current
phase,
and
that
would
be
the
true
matter
learning
I
think.
In
the
actual
sense,
what
we're
doing
in
memory
is
we're
learning
a
better
initialization.
A
It's
solving
the
problem
of.
I
only
have
a
few
samples,
and
what
do
I
do
right?
The
data
set
when
the
solvent
is
the
embedding
learning
is
solving
the
same
problem
and
the
memo
is
solving
the
same
problem.
So
all
this
is
three
different
approaches,
but
they
also
have
the
same
problem
and
you
could
directly
compare
like
mammal
to
matching
networks
to
relational
networks.
A
So
we're
going
to
benchmarks,
so
the
benchmark,
we're
using
here
right
now
at
momenta
is
omniglot.
I
think
karen
did
the
research
topic
on
this.
I
won't
talk
about
it
too
long,
but
basically
you
have
two
data
sets
background
and
evaluation
so
background.
It's
go.
It's
30
alphabets,
I
think
960
characters.
A
A
Exactly
yeah
exactly
where
is
all
the
quickness
of
learning
new
tasks
and
for
each
one
of
these
characters
you
have
20
samples
per
character.
So
this
is
what
it
looks
like
for
like
an
alphabet.
This
is
a
bengali.
This
is
sanskrit
alphabet,
so
each
one
of
these,
for
example,
this
letter
in
sanskrit.
We
have
20
samples
and-
and
here
is
what
categorizes
this-
how
machine
learning
sits
pure
shot.
So
you
can
see
it's
always
a
bunch
of
tasks,
but
only
a
few
samples
per
test,
but
it's
never
just
one
task.
A
B
Look
can
we
just
in
the
previous
slide
the
omniguard
thing
does
this
include,
like
you,
know
our
alphabet
and
would
distinguish
between
other
similar
alphabets,
I
mean
so
like
these.
Are
I
mean
I
don't
have
no
idea
bengali
and
sanskrit
share
anything
in
common
to
me.
Those
scribbles
right
but,
like
you
know,
if
I
look
at
oh
french
versus
german
they're,
pretty
similar,
you
know
some
few
accents.
So
does
this
this
data
set
include
those
kind
of
alphabets
too.
What
I
mean
what's
in
it.
B
B
Well,
yeah,
but
you
know,
there's
a
lot
of
alphabets
are
very,
very
similar
and
overlapped
slightly,
and
so
the
question
is:
are
they
trying
to
find
alphabets
they're
very
orthogonal
to
each
other?
Are
they
also,
including
like
oh
we're,
trying
to
distinguish
between
french
and
german,
which
are
all
nearly
identical,
because
there's
only
a
few
slight
differences,
you
know
a
few
extra
characters
here
or
there.
I
I'm
just
curious
of
that.
That's
that's
like
a
different
problem
than
trying
to
decide.
B
B
Yeah
they're
very
similar
it
so
that
it
just
it
just
I
don't
know
what
it
means
I
just
want
to
understand,
or
they
so
now
we
have
these
different
alphabets
that
are
that
these
characters
are
that
overlap.
They
exist
in
multiple.
They
exist
in
both
multiple
alphabets.
These
characters.
All
right!
That's
I
just
want
to
understand.
I
don't
know
what
it
means
yet,
but.
B
To
do
that's
clearly,
not
the
case
here
right
as
I
said,
it
would
be
easy
to
do
that.
If
I
took
three
alphabets,
you
know
the
english
letters
bengali
and
cyrillic.
Those
are
all
different,
really
different
and
and
there's
not
much
overlap
at
all.
But
if
we're
including
things
like
you
see
here,
then
there
is
a
lot
of
overlap
which
makes
the
task
quite
complex,
I'm
not
even
sure
what
it.
A
Is
anymore
yeah,
so
there's
no,
no
notion,
at
least
in
this
dataset
of
a
hierarchy
of
categories
of
languages
and
and
that's
a
problem,
and
I
think
even
worse,
is
that
in
this
training
paradigm
we
have
been
discussing,
we
usually
only
care
about
the
character
and
when
we
define
a
task,
we
randomly
sample
characters
from
different
alphabets
and
that
notion
of
you
know
if
alphabets
are
similar
or
even
if
the
characters
belong
to
the
same
alphabet.
It's
usually
not
taken
into
account
into
like
a
regular
fuchsia
training
paradigm.
C
B
B
A
B
Yeah,
but
we
also
don't
look
at
letters
in
isolation
very
much.
You
know
when
someone
gets
in
about.
Oh
I'm,
I'm
you
know,
reading
bengali
now
and
oh
I'm
going
to
be
reading
english
now.
There's
this
whole
contextual
thing,
that's
going
on
yeah
and-
and
you
can
very
quickly
look
at
even
just
a
single
word
and
say
well.
There
may
be
two
letters
in
here
that
are
the
same,
but
the
rest
of
it's
very
clear
what
it
is
so,
but
they're
not
doing
that.
A
It's
important
and
I'm
actually
going
to
get
to
that
point
a
bit
later,
and
so
the
second
data
sets
imagenet,
and
so
the
idea
here
is
mini.
Imagenet
is
just
100
classes
from
imagenet
which
are
taken
from
1000
classes
in
the
competition,
and
you
have
200
samples
available
per
class
and
the
images
are
downsized.
A
So
it's
a
it's
a
bit
easier
to
learn,
and
one
thing
here
is
imagine
that
I
don't
know
if
you
guys
know,
but
it's
actually
having
8
000
classes,
total
the
data
set
and
it's
taken
from
a
word
net.
It's
like
an
ontology
and
you
you
actually
have
in
the
word
night
you
have
like
as
these
relations,
so
you
can
tell
and
there's
actually
a
bag
and
the
classes
are
within
that
tag
and
you
you
have
the
distinction
of
animal
and
then
dog
and
then
dog
breed
and
so
on
so
forth.
A
But
in
this
competition
they
just
took
a
thousand
classes
out
of
the
total
image
in
that
data
set
and
those
are,
I
think,
those
are
all
leaf
nodes
right,
so
they
don't
have
relations
to
one
another
and
they
basically
disregard
that
relationship
there.
So
what
we
know
as
the
image
now,
what
we
usually
refer
to
as
imagine,
that
it's
those
thousand
classes
but
actually
imagenet,
is
a
much
larger
data
set
which
it's
based
on
this
ontology.
So
you
can
have
this
relationship
and
so
that
there
is
this.
A
B
A
B
A
Yeah,
it's
the
same
set
of
images.
It's
just
that
the
classes
they
sample
got
it
here.
There
are
something
from
lyft
nodes
and
here
the
same
thing
for
like
different
levels
in
the
tree.
You
can
open
it
like
that
now
another
one,
a
bit
older
with,
but
was
used
a
lot
in
the
past
is
the
birds
dataset
no
cub,
and
these
are
200
classes.
A
You
have
about
59
samples
per
class.
I
made
sizes
about
the
same
as
the
one
before,
and
one
thing
that's
a
bit
different
here
than
the
other
data
sets
is
that
those
are
mainly
balanced.
So
you
have
same
number
of
samples
per
class
which
makes
it
for
a
much
easier
problem
to
solve.
This
starts.
You
got
a
bit
unbalancing,
so
you
have.
This
is
a
histogram
and
you
have
some
some
images.
Some
classes.
You
only
have
20
inmates.
Some,
you
have
all
the
way
too,
but
the
average
is
60.
D
B
A
C
C
B
E
A
So
so
some
problems
with
those
data
sets
if
you're
thinking
about
you
know,
I
want
view
shop
learning
to
represent
human
learning.
So
one
is
that
these
are
completely
balanced
or
they're
fairly
balanced
a
data
set,
and
that's
not
how
real
life
looks
like.
So
it
should
vary
in
terms
of
number
of
classes
per
task,
and-
and
so
here,
when
we're
training,
we
separate
classes
into
tasks
right.
A
So
we're
training
like
five
classes
at
a
time,
and
we
call
this
a
task
and
it's
always
the
same
number
of
classes
and
it's
always
the
same
number
of
examples
per
class,
and
both
of
these
should
vary
both
in
terms
of
number
of
classes
per
test
and
the
number
of
examples
per
class.
Because
that's
that's
how
life
is
and
another
issue,
and
that's
actually
the
biggest
issue
here
is
that
the
test,
testing
and
training
dataset
are
drawn
from
the
same
task
distribution
and
that's
like
the
biggest
premise
in
all
those
algorithms
we
have.
A
So
we
are
only
ever
measuring
and
within
data
set
generalization
now,
when
we
actually
want
models
that
can
generalize
to
entire
new
entire
new
distributions
or
data
sets,
and
so
these
are
kind
of
the
two
of
many
issues,
especially
the
second
one.
Regarding
the
data
sets
we
have
and
the
training
pair
names
we
have
right
now.
A
What
we
usually
do,
we
should
learn
within
the
same
categories.
It
makes
sense
that
we
are
learning
dog
breeds,
we're
going
to
learn
different
dog
breeds
for
dogs
and
not
just
mix
a
dog
breed
with
a
cat
with
a
table
with
a
chair
that
doesn't
make
a
lot
of
sense
for
how
we
learn
and
that's
that's
one
of
the
issues
that
is
the
data
sets
don't
address.
Some
of
them
have
categories,
but
the
algorithms
that
deal
with
them.
They
don't
address
this.
A
So
one
one
proposal
that
came
out,
I
think
this
last
year
high
clear,
I
think,
that's
from
uft
with
google
and
open
ai.
Is
this
a
metadata
set
which
is
kind
of
inspiring
that
cvpr
10
task
fusion
benchmark
and
they
try
to
address
these
main
issues.
So
in
here
we
will
have.
We
have
actually
a
combination
of
10
data
sets,
and
so
tasks
can
be
composed
of
these
different
data
sets
it's
the
same
model
being
learned.
So
it's
not
a
continuous
learning
scenario.
It's
just
10
different
data
sets
mixed
it's
unbalanced.
A
A
It
like
I
said
it's
trained
on
different
data
sets.
So
these
are
slightly
different
distributions
and
you
you
can
you
have
things
here,
ranging
from
traffic
signs
to
birds,
two
characters,
so
there
are
very
different
distributions
and
you,
you
start
taking
into
account
the
hierarchy
as
well.
So,
for
example,
omniglot
you
sample
by
category
which
means
like
by
alphabet
and
an
image
now
to
sample
that
category.
Instead
of
just
randomly
sampling
characters
which
can
belong
to
any
alphabet
at
one
time,.
A
And
so
whatev,
what
I
wanted
to
show
here
is
the
data
set
names.
I
don't
want
to
go
into
the
results
here,
but
it's
just
comparing
different
few
short
learning
approaches
to
this
data
set,
so
k,
n
and
fine
tune
with
matching
ads
with
memo
with
relation
x.
So
these
are
all
comparisons,
and
then
this
is
a
version
that
this
paper
implements,
that
it's
a
combination
of
relationships
and
memos
and
bits
the
previous
benchmarks.
A
Every
drink
is
so
so
they
rank
their.
How
well
they've
done
in
that
test
compared
to
the
others,
and
this
is
the
average
rank.
So
so,
if
this
algorithm
was
like
the
best
one
in
all
ten
tasks,
there
would
be
like
one.
Every
drink
would
be
one
if
it
would
be
the
worst
ever
drink
would
be
seven,
because
there
are
seven
algorithms.
A
A
So
you
wanna,
you
want
to
learn
how
to
do
continue
learning,
instead
of
continue
through
short
learning,
instead
of
just
shot
learning
and
not
necessarily,
you
needed
the
few
shot
in
the
continual
learning
paradigm.
So,
when
you're
talking
about
oml
or
animo,
I
think
it's
more
of
a
hardware
restriction
and
kind
of
a
coincidence
that
they
set.
They
use
that
same
few
shot
paradigm.
But
if
you
only
want
to
learn
how
to
do
continue
learning,
you
could
even
take
the
first
shot
out
of
the
equation.
A
You
can
still
learn
it
the
same
way,
but
since
it's
very
compute
intensive,
they
ended
up
using
the
same
omniglot
and
the
same
data
sets
that
are
being
used
in
three
shot:
learning
those
are
kind
of
different
distinctions.
So,
right
now
we
are
doing
continuous,
shot
learning,
but
we
could
break
this
problem
and
think
about
just
few
shot,
learning
and
or
think
about
just
continue
learning.
Those
are
different.
B
C
A
C
B
C
Yeah,
this
is
a
meta
thing,
so
they're
setting
it
up
you
at
the
end
of
the
day.
You
want
a
network
that
can
learn
new
things
very
quickly,
yes
right,
and
so
you
want
and
how
you
get
to
that
network
shouldn't
matter
like
we,
you
know
with
humans.
You
know
we're
exploring
our
environment,
we
do
a
huge
amount
of
learning
and
then,
when
you
show
me
a
new
car,
I
can
learn
it
really
quickly
or
a
new
coffee
mug,
or
something
like
that
yeah
right.
B
C
B
C
D
B
C
C
A
A
A
A
A
There's
this
distinction-
and
there
are
people
now
thinking
about
this-
continual
few
shot
problem,
which
is
a
different
problem
than
fuchsia,
and
I
I
kind
of
wanted
to
go
over
this
and
highlight
because
we've
been
talking
about
fish,
shot
and
all,
and
I
think
we
should
keep
in
mind
other
approaches
to
the
same
problem
like
we
discussed
here.
Algorithms
approach
like
learning
optimizer
like
embedding
learning
approaches
and
those
that
might
give
us
some
insights
on
on
how
to
solve
it
or
even
what?
A
What
should
we
compare
our
approach
against
right
and
at
the
end
of
the
day,
but
I
actually
wanted
to
go
through.
So
this
is
one
thing,
but
also
the
benchmarks
and,
like
I
said
this
is
like
a
series
of
meetings
that
we
want
to
do
that
presents
some
of
the
benchmarks.
We
have
machine
learning,
so
you
can
see
what
we
have
now
and
then
you
can.
E
A
A
C
A
B
D
B
So
so
you
could
fit
some
sort
of
behavioral
characteristics,
part
of
a
behavioral
character's,
attentive
sort
of
component
into
these
benchmarks,
but
other
more
other
components
like.
Oh
I'm
rotate,
I'm
looking
at
something,
I'm
rotating
it
or
I'm
touching
something
and
moving
my
finger
that
doesn't
fit
into
these
things
at
all.
C
A
A
B
Cool
okay
thanks,
I
have
a.
I
have
a
few
comments
to
make
on
this,
just
a
a
great
presentation.
By
the
way
you
know
we
have
a
challenging
task
in
front
of
us
here
and
it's
similar
to
what
we've
been
doing
in
sparsity
right,
so
we're
trying
to
take
something.
We've
learned
about
the
brain,
sparse,
activation,
sparse
connectivity
and,
first
we're
applying
it
to
existing
convolutional
neural
networks,
existing
models
and
we're
trying
to
shoehorn
what
we've
learned
about
that
into
these
models
and
trying
to
do
trying
to
do
good.
B
But
ideally
we
want
to
go
away
from
that
right.
We
want
to
go
extremely
sparse
and
therefore
we
have
to
go
to
much
wider
models
if
we
will
right.
So
that's
that's
where
we
want
to
go,
but
we
have
to
initially
show
hey.
This
works
an
existing
model,
but
we
really
know
we
have
to
present
myself
the
same
issue
we
face
here
right,
we
would
say:
oh
maybe
we
can
apply
some
of
our.
B
What
we've
learned
about
fu,
shot
learning
to
these
existing
models,
but
maybe,
in
the
end
we
have
to
go
to
something
else
really
take
advantage
of
it.
Like
super
tight
suggested,
you
know
models
of
movement,
and
so
on
that
these
guys
don't
talk
about
at
all.
So
we
have
that
interesting
challenge
ahead
of
us
there
as
well,
and
I
I
also
want
to
point
out
that
we
know
at
least
two
of
the
major
components
of
the
solution
to
how
brains
do
this
and
I'm
just
going
to
encourage
when
we
think
about
how
we
go.
B
We
should
keep
those
two
components
in
mind
all
the
time,
and
I'm
repetitive,
you
all
know
you
should
know
this.
We
talked
about
it
a
lot.
One
is
the
issue
of
using
you
know,
dendritic
computation
as
a
way
of
of
solving
the
sort
of
catastrophic
learning
problem.
That
requires
us
to
settle
these
large
convolutional
neural
networks.
B
Very
slowly,
because
that's
that's
the
only
way
to
get
the
answer
and
the
other
is
this
whole
idea
of
reference
frames
and
attempt
and
attention
so
you're,
actually
building
models
of
in
a
reference
frame,
and
therefore
you
can.
All
you
have
to
learn
is
the
relative
positions
of
a
few
components
and
you
basically
got
them
that
amount,
and
so
those
are
the
two
things.
At
least
we
have
identified
so
far
and
we
just
have
to
make
sure
we're
moving
in
that
direction.
A
B
Yeah,
that's
the
first
thing
to
tackle.
I
think
I
think
niels
is
trying
to
work
on
the
second
part,
I'm
not
sure
exactly
what
he's
working
on,
but
we
need
to
get
there
eventually,
and
I
think
you
know
we're
starting
to
think
about
the
infrastructure
for
reference
frames
and
so
on.
But
then
that's
the
movement
with
matheran
streams
is
the
real
is
one
of
the
key
core
components
of
how
we
learn
quickly,
but
it
also
requires
the
computation
and
the
separation
out
of
problems
into
these
branches.
B
A
A
B
A
Thanks,
I
want
to
talk
about
earth
based
learning
and
I
think
you
remember
vivian's
presentation
and
it's
it's
it's
in
an
environment.
It's
getting
some
sort
of
input,
it's
learning
through
behavior
and
it's
learning
with
no
supervision.
So
I
think
it's
somewhat
closer
to
what
we're
looking
for,
and
maybe
we
can
get
some
insights
there.
I
just
want
to
go
next.
I
want
to
go
some
of
the
go
over
some
of
the
work
here,
but.
A
E
B
B
B
Well,
I
think
that's
in
the
human
brain,
it's
built
in
it's
just
it's
it's
automatic.
If
you
have
a
misprediction,
that's
built
in
it's
part,
it's
just
part
of
the
fabric
of
the
brain,
a
misprediction
says
attend
to
it
and
rebuilding
fix
that
part
of
the
model.
You
don't
really
have
a
choice
in
the
matter.
You
know
some
if
you're
just
touching
something
or
looking
at
something
and
there's
an
error,
you
just
you
it's
automatic,
you
tend
to
it
and
that's
that's
the
process
right.
There.
A
Yeah
in
reinforcement,
learning,
for
example,
you
can
think
of,
for
example,
you
have
a
robotic
arm
and
then
it
has
like
some
degrees
of
freedom
of
what
it
can
do
like
it
can
push.
It
can
grasp,
and
you
have
this
arm
and
you
have
like
a
bunch
of
thoughts
like
objects
and
what
curiosity-based
learning
that
don't
test
mean
for,
for
example,
is
you're
always
trying
to
explore
regions
of
the
state
action
space
and
you
haven't
explored
yet
so
the
arm
is
just
going
to
set
itself
a
bow
like.
A
Oh
I'm
just
going
to
try
and
kick
the
red
block
to
the
corner,
and
then
it's
just
going
to
try
to
learn
that
and
then
oh
I'm
just
going
to
try
and
grab
the
green
triangle,
and
then
it's
going
to
try
to
learn
that
and
the
way
it
defines
this.
It
looks
for
regions
of
the
state
action
space
that
it
has
low
exploration,
so
it
doesn't
know
yet
how
to
do
so.
It's
just
going
to
learn
just
you
know.
B
B
You
know,
especially
once
you've
reached,
you
know,
adulthood
in
some
sense
you
know
a
child
can
strike,
try,
sticking
its
finger
into
an
electrical
socket
and
do
various
other
things,
but
but
a
you
know,
once
we've
learned
about
the
world,
we
don't
we
don't
try
all
you
know.
I
don't
get
my
car
today
and
say
hey
what
would
happen
if
I
turn
the
wheel
all
the
way
to
the
left
and
hit
the
accelerator
and
you
know
try
to
shift
at
the
same
time.
B
A
C
B
B
I
I'm
just
pointing
out
that
curiosity
of
trying
unexplored
areas
of
sensory
motor
space
is
certainly
something
we
do,
but
it's
it's
certainly
something
we
wouldn't
want
to
do
with
abandon
and
it
needs
to
be
tempered,
and
you
know,
and
as
you
as
you
mature
it's
clearly
you
can
you
do
less
of
that.
You
know
the
older.
You
get
the
fewer
up,
the
fewer
you
know,
risks
you
take
them
on
general,
not
for
everybody
in
general.
B
B
Yeah
yeah
and
expectations
I
mean
clearly
what
we're
doing
is
we're
learning
what's
likely
to
happen
and
and
based
on
and
as
you
get
you
know,
as
you
get
older
more
and
more,
you
think,
like
I
think
I
know
what's
going
to
happen
here,
I'm
not
going
to
try
this.
I
mean
I
know,
what's
going
to
happen
when
you're
younger
you're
saying
I'm
not
sure
what.