►
From YouTube: NuPIC Office Hour (Attempt 2) - Feb 9, 2015
Description
No description was provided for this meeting.
If this is YOUR meeting, an easy way to fix this is to add a description to your video, wherever mtngs.io found it (probably YouTube).
A
Hello,
new
big:
this
is
the
new
big
office
hour
for
the
new
era.
Ninth,
we
had
some
technical
difficulties
with
the
other
hang
out,
so
it
just
left
it
open.
So
don't
join
that
one.
It
doesn't
work,
it's
not
even
running,
so
this
is
a
temp
2
and
it
seems
to
be
working
in
broadcasting
now
of
them.
So
I
am
a
map,
sailor
and
and
Friends
of
yoga
Jaden
suicide,
Jeff
and
Scott,
and
we're
going
to
be
answering
your
questions
from
the
new
community.
So
let
me
share
my
screen
here:
yeah
I
love
it
okay.
A
A
A
A
To
join
and
see,
what's
going
on
so
I
started
with
this
first
question
from
Michael
when
it
almost
ceased
nitin,
but
for
the
first
time
all
the
cells
in
the
column
column
active
one
sided
but
becomes
part
of
a
different
pattern.
Only
one
of
the
cells
will
be
predicted
for
active
state.
How
does
a
column
choose
their
cells
going
to
represent
the
expected,
but
the
based
on
connections
to
cells
of
different
columns
and
it
items
to
us
everything?
We
have
everything
Joey.
C
A
C
You
want
to
make
something:
that's
random
in
the
case
where
nothing
matches.
Otherwise
you
actually
won't
end
up
getting
a
hoarder
context,
so
it's
important
that
it
gives
some
event
in
between
a
week
match
and
completely
random.
We
look
for
the
least
recently
used
so
going
to
do
to
sort
of
fill
it
up.
Yet
it
encourage
good
description
and
if
there's
an
NB
tiebreaker.
B
A
A
C
A
A
Right
so
says:
CS,
so
I'm
just
putting
something
in
here
to
see
if
anybody
is
watching
from
in
the
meantime,
we'll
just
move
on
to
the
next
question,
because
it
looks
like
everything's
working
now.
Okay,
question
number
two
from
Michael
Roy
amps
he's
just
recently
signed
the
computers
license.
So
we
have
some
intent
to
help
us
work
on
this.
How
do
we
allow
new
learning
to
happen
while
guaranteeing
their
attention
a
valuable
hold?
Learning
a
new
big
Network?
Are
there
examples
of
networks
for
getting
older
learning?
B
C
C
In
general,
in
the
temporal
memory,
the
learning
is
contextual.
So
if
you
get
a
certain
particular
context
again,
then,
and
but
something
different
happens
in
the
future,
it'll
slowly
override
that
one.
But
if
you
don't
see
a
particular
context,
those
memories
will
stay
on
for
a
long
period
of
time.
No
well
I
know
what
is
it?
It's
a
there's.
B
A
couple
of
things
is
the
whole
idea
of
the
permanence
is
that
you
can
increase
the
longevity
of
a
synapse
beyond
point
where
become
very
effective.
I
said
I
get
you,
you
can
keep
increasing
the
permanence
up
that
one.
The
way
of
saying
this,
this
connection
needs
to
be
remembered
longer
than
one
that
have
gotten
one.
So
that's
the
basic
idea
and
it
doesn't
become
more
effective.
It
just
becomes
more
of
a
long
time,
so
it
takes
longer
to
forget
I.
Think
with
see
what
I
was
saying
is
that
capacity.
B
B
So
so
it
has
a
very
gradual
failure
or
gradual
degradation,
as
you
fill
up
and,
of
course,
things
that
are
more
program
lasts
longer,
but
as
soon
as
I
mentioned,
it's
biologically
possible.
That
seems
likely
actually
that
there
are
other
ways
that
make
the
learning
you
can
some
money
more
government
based
on
certain
factors,
but
we
model,
but
a
gentle
little
things
for
a
long
period
of
time
things
it
exposed
more
about
how
longer.
C
A
A
Tucker
point
is
missing
as
a
missing
critical
feature.
Only
bad
people
from
doing
great
things
he's
talking
about
priorities
in
networks
and
stuff.
Is
that
true?
It?
Yes?
Is
this
being
a
planet
over
time,
Hey,
look
at
the
road
map,
it
isn't
see
anything
about
that.
That's
true!
Is
there
a
different
road.
B
A
C
So
the
devel-
bet
is
focused
on
the
next
few
weeks
and
next
couple
months
for
new
pic
per
se.
Then
there's
the
new
pic
research
repository,
which
contains
all
of
our
research
code
and
all
the
tasks
that
we're
working
on
a
research
standpoint.
So
the
tempo
cooling
work
would
be
tracked
there
and
you
can
look
at
the
github
issues
there
and
the
wiki
page
there
and
we
don't
promise
to
keep
that
updated
regularly.
But
maybe
try
to
you
can
try
to
get
a
feel
of
what
we're
working
on
over
there.
C
C
B
B
B
C
We've
worked
through
the
lot
of
the
sensory
motor
and
Prince
and
Temple
Lane.
You
know
corner
cases
and
so
on.
Chay
nhanh
has
written
up
a
nice
linking
page,
exploring
kind
of
the
limits
of
sensorimotor
influence
and
tempo
cooling
like
how
many
different
things
you
can
pull
over
and
so
on
there.
So
there's
a
long
kind
of
write-up
of.
A
A
C
C
B
A
B
A
B
B
We're
not
trying
to
build
hierarchies,
we're
just
focusing
on
d,
I,
actually
think
the
jumper
bullying
what
is
easier
and
it's
a
simple
thing
and
we
sort
of
conceptually
put
got
set
to
a
point
website
at
the
moment.
Now
it's
required
to
do
it's
going
to
be
required
to
figure
out
how
you
know
all
the.
B
A
C
C
B
A
We're
going
to
continue
going
through
50
questions
right
now,
but
we'll
get
to
the
Q&A
questions.
Okay,
next
question:
okay,
all
I!
Think
all
of
these
next
three
several
questions
are
from
Eric,
lock,
locking
and
blocking
he's
the
guys
and
working
on
a
continuous
HTM
idea.
I've
been
kicking
around
connecting
with
him
a
little
bit
post
more
closely.
So.
A
A
B
A
B
What
we
did
we
don't
now
my
understanding
question
correct
it
made
some
funky
about
it.
Didn't
someone
think
if
I
got
the
question
wrong,
if
you're
trying
to
chasing
it,
sounds
like
stability
of
competitions
like
your
problem
with
representations
are
changing
and
you
can't
have
that
happen
quickly.
You
know
how
slowly
for
both
changes.
B
If
you
want
to
change,
we
did
have
a
couple
of
corner
cases
where
this
occurred
in
the
past,
and
one
was
the
case
where
you
had
a
very,
very
limited
amount
of
input
into
the
system,
so
that
the
system
actually
didn't
see
a
lot
of
different
patterns,
and
we
saw
a
few
times
and
you
know,
and
the
way
the
Susan
tries
to
work.
But
all
the
cells
are
trying
to
all
columns
with
Marcela
trying
it's
an
aggressive
learning
strategy
and
they
try
to
represent
something.
B
They
want
Jimmy
over
the
inputs
basins
with
into
a
nice
clean
divisions,
and
if
you
don't
have
enough
employ,
if
you
don't
hear
like
three
patterns
coming
into
that
they
can
fight
over.
So
that's
a
non-biological
problem
that
doesn't
occur
in
biology,
but
we
give
you
a
big
movement.
Some
special
code-
this
is
looking.
You
have
a
way
we
start
training.
A
new
who
has
a
new
spatial.
Pooler
is,
if
it
doesn't
seem
I,
don't
think
they've,
never
seen
enough
unique
things
coming
in
then
and
has
sort
of
Corky's
prevent.
B
So
we
had
an
idea
question,
which
is
the
aggressive,
be
able
to
sort
of
spread
out
a
lot
of
among
all
the
german
str
village
in
columns.
So
that
was
the
corner
case
that
didn't
lie
would
apply
to
anything
like
you
know,
like
vision
or
something
like
that.
I,
don't
there's
another
interpretation
of
that
I.
C
C
B
May
be
a
present
raindrop
trade-off
between
learning
rate
and
wow.
What
part
of
your
historical
data
stream
you
focus
on?
If
you
have
an
I
learning
rate
you'll,
be
the
system
will
basically
adapt
quickly
to
the
things
that
happened
pretty
recently
and
a
more
quickly
forgetting
that
from
the
past
you
have
a
slow
learning
rate,
it
basically
ignore
or
slot
it
take
a
much
longer
time
for
new
patterns
to
pay
speaking
play
a
role
so
I'm
slowly.
B
Noise
and
say
just
go
a
little
bit
of
the
soup
any
time
before
I
start
saying
it's
real
and
they're
from
stuff,
so
that
that's
not
a
problem.
It's
just
a
matter
of
where
you're
living
in
the
world
you
want
to
want
to
react
to
things
happen
quickly.
One
reactant
things
have
a
long
time
ago.
There's
no
Audrey
are.
We
know
where
you're
going
with
financial
tonight
and.
A
Idea
what
that
means,
and
so
that
each
configuration
cells
produces
the
eq
value.
It's
works
quite
well.
This
is
the
way
the
brand
of
does
it
for
the
evidence.
I
think
I
think
I
didn't
see
this
one,
so
super
tight
had
mentioned
before
we
start
this.
This
we
have
a
ceiling
classifier,
which
basically
takes
three
giselle
stores,
a
essentially
a
prediction
of.
B
A
Would
happen
next,
one
that
sells
active
and
then
the
cells
all
go
together.
A
new
me
use
that
to
come
up
with
the
prediction
for
our
Mary
steps
in
the
future,
so
I
think
what's
happening
here
is
that
the
q
value
store
are
essentially
a
predicted
master
award
for
each
possible
action
and
its
story
in
a
set
of
those
for
each
combination
of
active
cells.
I
think
that's
what
he's
describing
here
and
if
that's
the
case,
you
know
that
could
be
a
useful
way
to
do.
A
D
B
The
in
general
is
it's
difficult
to
take
concepts
that
would
you
ride
without
any
constant
consciousness
to
the
neuroscience
and
say
hey?
This
is
how
it
might
work
on
the
brandy
it
doesn't.
He
doesn't
work
that
well
that,
but
the
art
produce
really
does
as
Scouts
a
million
neurons
will
actually
implement
the
stuff,
and
then
something
like
these.
These
constant,
like
two
values,
could
influence
your
thinking
about
that.
But
it's
unlike
your
ears
out
yeah.
It
was
like
you
guys,
sort
of
this
snow.
A
B
B
A
B
A
Justified
this
community
seems
like
it
isn't
such
a
big
deal,
that
the
amount
of
every
need
goes
up
linearly,
but
he
seems
to
think
otherwise.
The
other
option,
which
I
assume
you've
seen
us,
use
all
possible
combinations
of
this
to
represents
the
cells.
This
is
how
intense
HTM
does
it.
This
is
Eric's
project,
but
I
do
not
have
theoretical
justification
yet
only
experimental
when.
B
B
Some
people
very
naively
think,
oh
well,
my
30
cells
in
the
column,
I
can
remember
elements
and
our
three
different
representations,
which
is
not
true
all
as
we
make
food
many
times,
is
that
your
voice,
as
a
set
of
column
representations
in
a
set
of
cells
that
the
number
of
possible
combinations
of
oneself
or
column
out
of
n
number
of
columns
it's
astronomically
large.
So
we
don't
have
actually
any
representational
tissue
at
all
with
our
reputations
and
copters.
Not.
B
It
is
astronomically
large
number
so
that
you
don't
in
memory
what
we
doesn't
go
up,
linear
all.
As
we
said
earlier,
we
work
for
a
fixed
amount
of
memory,
County
we
were
living
to
heaven
and
it's
quite
quite
large
capacity
in
this
regard.
In
so
this
issue,
that's
that's
what
jump
engines
thinking
he
misunderstood
how
those
cells
and
columns.
C
C
B
A
B
B
B
A
B
With
the
Ottawa
jamala
times
we
kept,
who
knows
my
head,
yeah
I
think
guys
were
doing
something
in
neural
network
architectures
more
than
us
for
us.
So
what
I
don't
know
that
I
in
trouble?
I,
don't
know
what
anything,
but
it
seems
unlikely
that
that
feature
is
an
absolute
necessity,
so
it
doesn't
like
I,
don't
focus
instead,
that's
the
beauty
of
the
secret
to
jump
about
me.
It
has
a
beautiful
way
of
doing
that.
Just
right,
I
already
come
up
with
nice
way.
A
A
Worse,
it's
also
a
couple
people
chatting,
but
no
one
asking
questions
so
I
think
that
is
it
I,
so
that
was
a
new
big
office.
Half
our
friend
Aaron,
the
morganite,
because
I
have
any
more
questions
we
might
put
off
office
hours
for.