►
From YouTube: HTM Chat / HTM Forum Q&A - May 6, 2019
Description
Broadcasted live on Twitch -- Watch live at https://www.twitch.tv/rhyolight_
B
A
Is
the
yeah
I
just
threw
a
link
into
chat?
This
is
the
place
to
be
if
you're,
into
HTM
hierarchal,
it's
a
forum,
memory
or
just
brain
inspired,
intelligence,
algorithms,
but
we're
talking
like
really
brain.
Inspired
like
we
want
to
understand
how
it
works
in
the
brain
and
be
able
to
like
teach
people
how
it
works,
not
we're
not
trying
to
come
up
with
a
formula
or
I.
Don't
know
how
to
say
it.
We.
A
We
don't
just
want
to
create
solutions
based
on
a
loose
interpretation
of
how
neurons
in
the
brain
work.
We
want
to
really
understand
the
core
of
intelligence.
So,
if
you're
interested
in
that
here's,
the
place
to
go,
I
always
saw
this
on
Twitter.
This
is
interesting,
so
bit
kings,
one
of
our
most
active
and
popular
members
on
the
forum
and
he's
online
right
now,
I
know
so.
This
pattern
induces
predictable
and
repeatable
activity
deep
in
the
brain.
A
A
So
it's
all
visuals,
oh
okay,
so
hierarchy
as
deep
is
but
well.
The
question
is:
are
they
I
would
assume
they
would
they
would
induce
predictable
and
repeatable
activity
everywhere
in
the
cortex
v1?
The
visual
cortex
excuse
me
in
the
visual
cortex
v1
v2
v3
v4
at
least
that's
what
the
thousand
brains
theory
I
think
would
suggest.
Is
that
what
they
found.
A
A
A
A
A
There's
like
a
there's,
a
little
bit
of
a
feedback
loop
right
now
between
all
the
stuff.
We
have
learned
from
machine
learning,
because
that
is
I
mean
you
have
to
remember
that
all
originated
from
neuroscience.
The
the
the
point
neuron
models
that
we
have
all
totally
originated
in
neuroscience
like
I'm
always
complaining
like
this
is
too
simple.
It's
too
simple.
We're
hit
a
brick
wall.
We're
not
gonna
go
any
further.
A
A
A
A
A
A
Think
so,
just
talking
about
the
benefits
of
competition
when
it
comes
to
heavy
and
learning
so
as
an
example
and
is
in
HTM,
there's
a
mini
column,
competition
during
spatial
pool
linked
to
we
call
it
a
competition
in
amongst
neighborhoods
as
many
columns
to
decide
which,
which
mini
column
will
represent
the
current
input.
So
that's
how
we
sort
of
talk
about
that.
A
So
we
try,
we
apply.
What
do
we
call
it
boosting
John?
So
that's
sort
of
our
answer
to
the
problems
that
arise
with
heavy
learning,
not
the
competition,
even
with
it
in
the
mini
column,
competition.
You
can
have
the
same
two
problems
here
and
it's
not
just
about
the
threshold
that
said
in
their
cells
that
just
never
fire.
They
end
up
not
not
being
selected
for
anything
in
the
input
and
then
there's
cells
that
represent
too
much
and
and
so
that
keeps
the
cells.
A
So
this
is
that's
what
boosting
does
even-even
while
we're
having
these
local
competitions
we
still
applied,
boosting,
so
that
which
is
a
basically
to
look
at
the
active
duty
cycles
of
all
the
mini
columns
like
how
much,
if
they've
been
active
over
time,
and
we
squash
down
the
ones
and
and
reinforce
that
new
meaning
is
sort
of
assigned
to
columns
that
are
not
being
very
active.
And
there
is
evidence.
A
Is
happening
on
the
brain
yeah,
they
move
the
activity
to
keep
the
learning
threshold
of
the
outfit.
So
there's
there's
different
ways
that
you
might
try
and
do
this.
That
might
work
better
and
the
thing
there's
a
lot
of
things.
I
mean
this.
This
whole
world
that
we're
in
where
that
HTM
is
a
part
of
you
know
the
brain
biologically
inspired
cortical
inspired
world
there's
a
whole
lot
of
different
doors
that
are
open
right
now,
research
doors
like
what
I
could
do
this.
This
way,
you
could
fund
a
whole
study
on
something
like
this.
A
A
A
A
Would
think
you'd
want
well
I
guess
after
they've
learned,
perhaps
that's
the
case
unless
these
are
the
remaining?
These
are.
These
are
the
sort
of
on
the
fringes,
neurons
or
synapses.
Excuse
me,
they're
still
in
a
state
of
decay
or
growth,
but
I'm
not
I,
don't
quite
get
this,
maybe
I'll.
Maybe
y'all
and
y'all
know
that
figure
five.
A
A
So,
there's
what
you
could
do,
I'm
talking
about
current
AI,
you
could
take.
You
could
use
some
of
the
NLP
tools
that
they
have
today
to
learn
like.
First,
you
have
to
find
a
bunch
of
data,
so
you
need
to
find
a
bunch
of
words
or
text
or
something
that
that
a
character
might
have
said
or
used
or
something
like
that
or
you
need
to
create
a
lot
of
that
data.
A
I'm,
not
sure
if
this
is
gonna
work
and
then
it's
basically,
you
trade,
your
bottle
on
all
the
text
and
give
whatever
the
user
is
saying,
as
some
type
of
cube,
I
don't
know,
but
there's
certainly
things
you
can
do
in
the
world
of
like
text
generation.
There's
a
lot
of
big
development
going
on
there,
so
you
could
make
them
feel
more
realistic
by
making
that
there
chat,
sort
of
AI
powered,
perhaps
but
they're
not
gonna,
be
really
smart.
I
mean
they're,
not
really
gonna,
understand
anything.
It's
just
going
to
be
probable
as
probabilities.
A
There's
a
good
probability.
This
is
a
rewarding
thing
to
say,
depends
on
how
you
train
your
NPC
but
another.
So
so,
as
far
as
like
HTM
wise
I
think
perhaps
the
the
better
thing
to
focus
on
for,
if
you
want
to
try
and
use
HTM
like
brain
based,
stuff
is
location
and
space
and
movement.
So
we
do
have
a
pretty
good
idea
of
it
from
ways
to
map
space
and
keep
memory
of
where
you
have
been
in
space
and
and
the
paths
you
have
taken
through
that
space.
A
So
you
could
potentially
train
an
AI
to
navigate
more
efficiently
with
HTM
through
through
the
space
and
once
you've
done
that
you
can
basically
give
that
map
to
any
NB
C,
of
course,
but
but
honestly,
that's
not
a
hard
problem
anyway.
So
I
don't
know
if
it's
gonna
give
you
a
lot
of
benefit,
because
it's
like
brute
force
navigation
is
pretty
easy
and
they've
already
solved
it,
but
a
lot
of
the
Maisy
stuff
with
math
just
math.
You
know
it's
like
all.
A
A
B
A
A
Here
too,
because
they're
interested
in
brains
or
algorithms
or
math
or
both
you're
gonna,
be
let
down,
if
you're
here,
to
learn
about
math
as
I'm,
not
a
mathematician
reinforcement,
learning
in
HTML
Gotham's
first
post.
This
is
a
person
as
a
master
student
to
research,
a
deep
learning,
mostly
in
RL.
A
B
A
A
An
introduction
to
concepts
I
want
to
see:
where
is
the
deep
Learning
Network
in
this
system
well
and
I,
think
they're
in
like
encoders
and
decoders
right,
so
there's
there's
some
type,
but
no
that
was
for
transformers
all
right.
Let
me
start
from
the
so
so
somehow
you
take
state
and-
and
you
have
like
this
decision-making
module
and
you
have
to
incorporate
reward
and
feedback
into
that.
A
Markov
stage
reinforcement
learning
agent,
so
this
this
policy
function,
the
agent's
behavior
function
is
a
map
from
state
to
action,
so
I
think
we
need
a
a
stochastic
property
right.
How
is
that
don't?
Look
at
me
negatively
for
looking
this
up.
Stochastic
is
randomly
determined,
okay,
all
right
so
random
probability.
Distribution
deterministic
is
not
right,
not
random.
So.
A
B
A
Sort
of
established
as
a
spacial
spatial
search,
a
spatial
problem
right
from
what
it
seems
like
if
you're
trying
to
create
all
of
the
reinforcement,
learning
like
a
I,
open,
AI,
gym
or
whatever
they're
set
up
to
solve
this
in
a
spatial
way,
but
you
need
a
model
that
can
change
at
every
time.
Step!
That's
the
thing!
You
need
a
model
that
can
change.
I,
don't
know
if
reinforcement
learning
systems
will.
B
A
Things
that
are
familiar,
or
maybe
even
perhaps,
learning
something
that
are
leaning
towards
observing
something
that
you
already
know
a
little
there's
a
lot
of
things
that
you
do
so
the
problem
is
that
all
the
RL
systems
today
are
still
set
up
as
a
as
spatial
processing
problems
and
HTM
model
needs
the
needs
to
change
at
every
time.
Stone.
A
A
A
A
A
A
A
A
Alright,
the
mechanism
behind
lucid
dream
before
you
can
think
about
the
mechanism
behind
lucid
dreaming.
Is
you
have
to
sort
of
explain
the
mechanism
behind
dreaming
I,
don't
know
that
science
I,
don't
think
science
has
answered
that
that
I
I
mean
I
you
can
I
have
some
ideas
but
they're
just
ideas.
A
A
A
A
Was
this
other
one
visualizing
a
neural
machine
translation
model?
This
is
seek
to
seek
with
intention.
I
wanted
to
understand
how
transformer
works-
and
this
said
in
the
previous
post,
with
a
good
attention
and
attention
is
like
a
mechanism,
a
neural
training
mechanism.
I
was
trying
to
understand
this.
A
A
And
I
don't
know
why
I
just
was
curious,
because
I
I
see
them
I
see
them
in
the
text
that
I'm
reading,
sometimes
when
I
want
to
figure
out
how
they
work.
Because
these
terms,
like
especially
encoders
and
decoders,
you
know
they're
related
to
their
terms
that
we
use
in
HTM,
but
they
mean
sort
of
different
things
and
I
want
to
know
when
I'm
talking
to
somebody
about
this,
what
how
my
terms
are
overlapping
with
the
current
terms,
I
James
stammers.
A
A
A
A
A
A
A
A
For
twitch,
anyway,
build
HTM
systems
at
least
one
day
a
week,
and
then
I
need
to
have
like
a
Python
3
I
want
to
get
HTM
and
Python
3
working
properly
well
and
I
want
to
help
the
community
get
to
a
point
where
people
can
pip
install
some
HTM
I
want
to
get
them
so
that
the
community
fork.
This
can
be
pip,
install
a
ball
from
Python
3
and
once
we
get
that
I'll
probably
stop
doing
that,
because
we
need.
A
That
stuff
I'm
just
basically
getting
ready
for
a
research
of
meeting
at
10:00
and
I'm,
trying
to
like
wrap
up
doing
a
bunch
of
little
things
until
that
until
that
happens,
and
the
research
meeting
will
probably
go
on
for
a
couple
hours,
one
of
the
things
I
wanted
to
prep
for
the
research
meeting.
Was
this
paper
review
because
Supachai
brought
it
up
and
it
might
be
something
we
talked
about.
So
let
me
do
this
real
quick.
We
find
this
Lacey
I,
don't
mind.
You
guys
see
this
I've
got
this
giant.
B
A
A
A
A
B
A
Deadpool,
the
movie
he
uses
Adobe
Premiere
and
he
made
the
template
he
used
for
the
whole
project,
like
his
project
template
for
premiere
available
on
the
internet,
so
I
downloaded
that
sort
of
customized
it
a
bit.
Then
this
worked
out
pretty
well,
it's
a
if
it'll
ever
open
it
will
it's
not
opening
it's
taking
forever
to
open.
So
so
what
I'm
going
to
do
is
as
soon
as
this
opens
for
basic.
It's
like.
B
A
That's
just
something:
I
can
do
in
10
minutes,
because
once
I
sit
down,
then
I'll
be
ready
and
I
can
sit
down
at
premiere
and
it's
been,
you
know
an
hour
doing
an
editing
session
because
I,
don't
that's
something
like
programming.
Editing
is
sort
of
similar.
You
want
to
get
your
headphones
on.
You
want
to
get
the
zone.
You
know
it's
it's
hard,
there's
a
lot
of
different
like
if
you
want
to
do
something
right.
A
This
is
ridiculous.
It's
still
loading!
Oh
here
we
go
that
took
forever.
Okay,
this
project
was
saved
from
a
previous
version.
This
is
what
we're
looking
at
here
and
it
wants
me
to
save
it
as
that's
fine,
whatever
okay.
So
that's.
Why
I
did
that,
because
I
I
want
that
version,
so
I'm
going
to
shut
this
down,
remove
this
whole
one,
and
so
this
just
updated
the
template
every
time.
There's
like
there's
a
breaking
change
every
time
you
come
date,
your
project
and
you
have
to
save
it
as
a
different
version
of
your
project.
A
A
A
A
Okay-
and
this
is
gonna-
be
in
sound
production,
sound
all
right,
little
anal
I
know,
but
this
is
important
once
you
start
getting
a
complicated
project
and
it's
hard
for
me
to
tell
what's
going
to
get
complicated
so
then
you
have
to
drag
it
over
here
like
this
and
drag
it
into
premiere.
So
now,
I've
got
this.
B
A
A
Mmm-Hmm
I
don't
understand
this.
If
this
pisses
me
off
I've
been
using
this
recorder
for
ages
it
the
game
should
have
been
higher
I.
Don't
understand
why
the
game
was
an
iron
anyway.
I'm
not
gonna,
deal
with
it
now.
I
just
want
to
get
the
things
in
the
right
places,
so
the
other
place
was
footage
camera
a
wide,
we're
gonna
move
that
into
sequences
known.
A
All
right
so
now
we're
going
to
create
a
sequence
called
so
there's
already
a
sequence
called
selects
and
we'll.
So
we
use
that
so
we're
going
to
drag
the
main
movie
over
there
keep
the
settings
of
the
movie
and
we'll
have
to
change
the
settings
of
that
and
we're
going
to
drag
the
audio
also
into
here
and
we're
going
to
sync.
It.
A
A
A
A
A
A
A
Yeah
so
I
hope
it
I
haven't
read
the
whole
thing
yet,
but
I
took
a
perusal
yesterday
and
I
was
like
this
is
interesting,
so
they're
talking
about
cortical
columns
and
we're
always
looking
for
things
that
that
are
in
the
area
that
we
are
studying
and
we
were
studying,
cortical
columns
and
many
columns
and,
and
so
anything
that
comes
out
of
the
experimental
neuroscience
community
is
immediately
relevant
there
just
to
the
Pitkin
validation.
A
B
A
A
A
So
I
read
through
this,
so
let's
go
to
the
graphics,
cuz,
there's
I,
don't
completely
understand
calcium
imaging,
but
I,
but
I
know
it's
a
way
that
you
can
get
I.
Think
voltage
unless
I
think
that's
what
we're
showing
here
is
actual
voltage.
Spikes
of
an
individual
cell,
three
individual
cells
in
one
layer,
this
being
l5
and
the
same
thing
over
here,
so
we're
and
more
than
just
three
I
think
there's
available
here
is
more
than
just
three
I
think
this
is
just
showing
an
example.
A
A
A
A
A
A
Well,
at
least
you
have
a
chance
to
look
at
marks
already
seeing
it
over
the
the
tone
tone
not
to
be
as
I'm
interested
in
the
in
the
Tanaka
P
here
and
I
haven't
fully
decoded
this
chart.
Yet.
B
A
Just
don't
fire
and
during
your
experiment
they
don't
fire.
Okay,
so
report
on
those
put
that
in
your
paper,
how
many,
how
many
of
them
that
you
recorded
didn't
fire,
because
that's
important
context,
it's
interesting
to
know
that
those
neurons
were
not
involved
in
that
particular
processing,
whatever
it
was
processing
at
that
time,
they're
involved
in.
A
A
A
A
A
So
I'm
trying
to
work
out
a
schedule
honestly
I'm
gonna.
This
is
work.
I
go
through
a
lot
of
white
words,
so
this
is
Monday.
This
works
out
well,
so
I'm
sort
of
getting
in
to
have
it
just
coming
in
the
office.
Turning
on
twitch
the
live
stream
and
just
having
a
casual
form,
QA
and
chat
with
people,
so
I'm
calling
that
HTM
chat
and.
B
A
A
B
A
Pretty
good,
this
was
generally
pretty
good
this
one
after
a
break
for
lunch
and
come
back
at
1:00,
because
this
is
a
1:00
p.m.
hardly
anybody
is,
is
watching
at
that
point,
so
I'm
thinking
about
moving
it
even
maybe
even
moving
the
one
today
we're
going
to
talk
about
semantic
fingerprints,
it
seems
like
the
morning
shows
are
better
so
maybe
I'll
just
move
this
Tuesday.
A
A
A
A
Okay,
I'm
I'm
gonna
be
connecting
to
the
office
for
a
moment
because
I
am
gonna,
have
a
stand-up
meeting
in
about
20
minutes.
So
you
guys
won't
be
able
to
see
this,
but.
A
A
B
A
A
B
A
B
B
A
A
A
A
A
B
B
A
And
I'm
still
sort
of
trying
to
decide
whether
I
want
to
move
this
I'm
gonna
just
keep
it
today
and
just
plan
up
next
week.
So
I'll
keep
this
one
where
it
is
today
and
plan
on
next
week
to
change
it
next
week.
Why
don't
we
talk
about
hopfield
network
soar,
Boltzmann
machines
for
our
AI
chat?
How
about
that
mark
I'm
gonna
be
around
for,
but
we'll
do
it
on
Tuesday
that,
like
I,
don't
know
9
or
10
Pacific,
which
maybe
it's
not
a
good
time.
I,
don't
know.
A
A
A
A
A
A
A
A
B
A
A
A
Okay
back
to
events
building
HTM
systems.
The
last
thing
we
did
was
scaffold.
Okay,
so
now
we're
just
gonna
be
adding
diagram.
So
this
I
am
debating
whether
to
do
building
HTM
systems
on
Friday
or
Thursday,
because
Friday
I
get
more
views.
It
seems
like
Friday,
is
more
popular.
Ten
minutes
till
my
stand-up
meeting,
okay,
snooze
snooze
and
lose.
A
A
Okay!
Well,
you
guys
are
stared
at
me,
I'm
just
going
to
make
sure
that
all
of
my
engineering
tasks
are
updated,
updated
because
believe
it
or
not.
I
am
still
an
engineer.
I
still
do
some
engineering
things
at
Numenta.
Let's
see
yeah
I
got
that.