►
From YouTube: HTM Chat - Forum Q&A - May 2 2019
Description
Broadcasted live on Twitch -- Watch live at https://www.twitch.tv/rhyolight_
A
A
That's
such
a
hard
question
to
answer.
It
seems
so
simple,
but
you
have
to
know
about
data
and
patterns
of
data,
especially
sequential
patterns
and
data.
I
mean
imagine,
lots
of
waves
sort
of
occurring
over
time.
Aggregation
is
sort
of
a
way
to
sample
reality
so
so
that
you
can
continue
to
keep
the
patterns
represented
in
whatever
the
sample
is
so
to
know
how
to
aggregate
your
data.
You,
you
have
to
understand
your
data.
It's
different
for
each
data
set
I
would
say
so.
I'm
gonna
leave
it
at
that.
A
I,
don't
know,
I,
don't
have
a
good
answer
for
that,
but
I'll
mark
it
and
I'll
post
that
video.
It's
my
lazy
way
of
answering
questions,
hello.
It's
on
bento,
we've
interacted
quite
a
bit
over
the
years.
I
have
just
found
an
interesting
paper
at
a
nice
2019
conference
author
uses
fast
ret
and
the
key
points
descriptions
for
images
fast,
retina,
key
points,
I,
don't
think
I've
heard
of
that
bunny
has
sparse.
Just
Hey
look
at
that:
it's
like
directly
related
to
HTM
binding
of
sparse
distributed
representations
in
hierarchical
temporal
memory.
A
The
author
was
what
was
his
name
Luke,
let's
see
if
we
can
find
it.
No
I
was
for
the
oud
you,
okay,
so
there's
names
not
there,
but
they
might
not
have
it.
They
probably
don't
have
them
all
listed
out
like
the
workshops.
It's
just
organizing
teams
and
the
main
speakers,
but
so
that's
cool
announcement
agenda,
tutorials
demos.
A
I'm
not
paying
attention
to
whether
this
is
an
act.
Hey
hey
mark
I
bet,
oh
you're,
probably
not
there
anymore,
whoever's!
Listening!
If
you
know
this,
I've
got
a
record.
I
have
a
record
that
has
two
copies
of
the
same
basic
out.
Basically,
album,
hey
Mark,
you
might
know
this
because
you're
sort
of
an
audio
file
and
you're
an
older
generation
that
would
appreciate
appreciates
this
form
of
media.
Yes,
the
king
is
in
I've
got
one
mono
and
one
stereo
record
of
the
same.
Album
are
those
because
this
was
recording
like
50-something
six.
A
Well,
it
says:
copyright,
1972,
but
I
know
these
recordings
came
from
earlier,
because
they're
all
pre-recorded.
This
is
like
a
anyway
ones,
mono
ones,
stereo
were
those
actually
too
physical,
physically
different
recordings
in
the
in
the
time
that
they
were
recording,
different
tracks
for
mono
and
stereo,
because
I'm
sitting
here
listening
to
the
mono
version,
trying
to
compare
it
to
my
mental
model
of
the
stereo
version
that
I
hear
all
the
time
trying
to
figure
out
if
it's
actually
different.
A
I'll,
let
me
look
through
the
presentation,
HTM
structure,
region.
Okay,
we've
got
to
be
careful
here.
We
want
to
say
mini
column,
I
think
yeah.
Let's
call
these
mini
columns,
but
we
were
bad
at
this.
In
the
beginning,
we
were
bad
calling
this.
Oh,
the
record,
lay
the
worst.
Oh
I'm,
interesting,
okay,
so
that's
cool,
that's
obviously
that's
the
value
of
having
an
extra
record,
because
it's
a
totally
different
recording
of
the
artist.
A
So
that's
neat
hello,
James
I
was
I
was
just
chatting
about
vinyl
records
with
bit
King
and
going
over
our
forum,
htm'
forum,
computers
aren't
real.
Perhaps
now
what
is
real.
Nothing
is
real
if
you
want
to
learn
what
we're
talking
about.
It's
hierarchical
to
form
memory,
there's
a
link
in
chat
there
and
explain
sort
of
the
concept,
so
they
were
cut
separate
least
two
different
masters
for
pressing.
They
may
be
the
same.
Recording
oh
I
figured
that
the
interesting
thing
about
that
for
someone.
A
A
A
Yeah
Mysterio
was
new.
It
was
a
new
feature.
Okay,
I
want
to
look
through
this
special
cooler
structure,
so
these
are
many
columns.
This
is
their
like
potential
pool,
/
receptive
field.
Each
seller
see
the
same
feed-forward
input
and
put
flows
through
proximal
synapses.
You
know,
there's
approximate
connections,
heavy
and
learning
online
learning,
it's
a
good
summary
sort
of
breaking
out
what
an
SDR
is
as
representation
it
represented
in
spatial
for
this
nice
pictures,
I
like
it
when
people
make
their
own
pictures
so
until
they
have
an
understanding
of
things,
non-local
is
distributed.
A
Non-Local
asturian
such
an
interesting
way,
to
put
it
learns
to
have
similar
cuts,
similar
outputs
similarity
can
be
computed
with
a
dot
product
and
then
the
essential
algorithm,
explaining
overlap
comparisons
and
then
inhibition
I,
don't
think
I've
seen
these
out
I
think
he's
these
algorithms.
These
aren't
coming
from
our
documents.
I,
don't
think
so.
That's
interesting
too
I
mean,
of
course,
there's
lots
of.
We
had
this
all
came
from
intuition
of
looking
at
how
neurons
were
working
and
trying
to
figure
out
what
they
were
doing.
So
we
didn't
start
with
the
math.
A
You
know
you've
got
to
go,
find
the
math
that
explains
the
biology.
So
it's
interesting
to
see
people
finding
the
math
for
it
right.
That's
cool
and
I'm,
not
gonna
verify
whether
this
is
right,
I'm
not
a
mathematician,
but
this
looks
somewhat
feeling
of
the
sets
like
I
said:
I'm,
not
a
mathematician,
but.
A
A
A
This
must
be
his
motivation,
perhaps
combined
multimodal
data
without
increasing
dimensionality.
Definitely
that's.
This
is
HTM
as
a
way
to
do
that
for
sure.
We're
gonna
check
out
the
asymmetric
confidence
thread.
I
have
some
things
to
look
at
that
may
be
wrong.
Yeah
I'll
pop
over
there
about
10,
more
minutes,
I,
don't
he's
talking
about
the
binding
problem,
vector
symbolic,
architectures,
vector
operations,
difference
different
implementations
of
vs
A's.
A
I
am
not
familiar
with
these
sparse,
distributed,
sparse
pioneer,
distributed
representations,
I.
Think
I.
Thanks
for
the
follow
James
interesting
enough
that
you
could
stick
around
for
the
next
one.
You
know
I'm
I
have
a
schedule
if
you
want
to
look
at
it
and
I
try
and
put
events
up
on
my
vets
page.
So
if
you're
on
Twitch
look
down
and
at
my
schedule,
that's
usually
right
but
there's
an
event-
a
link
to
an
events
page
that
I've
trying
to
keep
totally
updated.
A
Okay,
so
this
is
cool
I,
don't
binding
and
unbinding
I'm
gonna
I'm
gonna
skip
some
of
those
back.
Simoni
sparks
STRs
use
a
local
vape
on
a
mission
official
or
vast
retina
keypoint.
So
this
is
what
I
have
not
frickin
coding.
I've
never
heard
of
this
before
faster
retina
key
point,
so
we're
looking
at
like
motile
multimodal
data
or
no
difference
of
gaussians
I.
Guess
it's
something.
I
have
to
look
into.
Is
the
free,
fast
retina,
a
key
point
vast
retina?
A
A
Boy,
okay,
so
it's
like
basically
gonna
have
to
read
a
paper.
I'm,
not
gonna,
read
the
paper,
because
if
it's
interesting,
if
somebody's
interested
in
the
fast
threat
in
a
free
point,
you
think
it's
a
valid
thing
to
relate
to
HTM,
which
this
guy
obviously
does,
but
basically
I'm
gonna.
Ask
someone
to
explain
this
to
me
on
the
floor.
A
A
You
did
a
scan
of
the
paper.
Okay,
cool
cluster
analysis,
finding
experiments,
I
love,
seeing
this
stuff,
though,
even
though
I
don't
understand
what
what
it's
doing
I
mean
the
thing.
The
ideas
behind
HTM
are
gonna,
be
behind
a
lot
of
things.
You
know
of
understanding
more
about
how
well
you
can
do
a
lot
of
stuff
with
it
I
think.
So.
That's
very
cool,
that's
great!
Really,
that
is
an
is
really
interesting
for
sure.
A
A
Why
does
it
continue
to
go?
Why
does
it
go
up
and
down
with
the
data
and
I?
Don't
think
anybody
really
understands
why
my
my
initial
intuition
was
because
there's
because
there's
no
reset,
so
it
never
knows
what
psyche,
what
weather
is
it
in
the
one
period
sequence
or
the
two
period
sequence
or
the
three
period
sequence,
because
it's
just
building
up
more
and
more
and
more
and
more
and
every
time
it
ceases,
and
this
is
the
value
again.
A
It
can
branch
off
and
create,
say:
oh
I,
don't
this
might
be
a
new
sequence
or
the
old
sequence,
so
this
one
might
be
a
new
sequence.
Oh
this
one
might
be
a
new
sequence.
It
never
recognizes
that
this
is
just
one
sequence
repeated
over
and
over
again:
okay,
that's
the
that's!
The
core
problem,
I
think
that's
probably
causing
this
fluctuation.
A
A
So
it
makes
sense
if
the
confidence
was
consistent,
always
good
are
always
bad
or
always
jittery,
so
I,
don't
know
why
what
causes
it
to
change
characteristics
over
time,
that's
kind
of
weird.
But
that
being
said
again,
this
is
such
a
dynamical
system.
It's
hard
to
anticipate
what
it's
going
to
do
honestly,.
A
Last
two
graphs
unchanged
of
the
finite
values,
how
many
I'm
trying
to
get
to
I
can't
remember
all
the
context
and
I
want
to
thank
because
originally
it
was
fill
and
then
recently
that
we
didn't
really
find
an
explanation
for
this
as
far
as
I
could
tell,
but
then
pooling
if
I'm
saying
that
right
looks
like
so
I'm
interested
pulling.
A
If
it's
just
if
it's
just
simple
to
do
that,
like
like
I
asked
you,
maybe
change
the
jitter
to
a
different
value,
and
you
did
seem
to
do
that
pretty
easily,
but
I
still
like
what
I
was
saying
is
I.
Don't
understand
why
anomaly
score
goes
up
and
down
with
the
wave
unless
it
has
something
to
do
with
it
being
unsure
about
well
yeah
anyway,
let's
read
through
what
Vic
King
said.
So
we
see
stuff
like
this
when
doing
integer
math-
and
you
were
at
an
article
about
this
too-
that
I
thought
I
figured.
A
You
would
link
here
anyway
on
imp
use,
I,
don't
know
the
code
and
totally
uninterested
in
digging
through
the
nested
dopey
stuff.
Yeah
I
don't
want
to
you
know
it's
not
so
bad
in
the
spatial
Pooler.
Most
of
that
logic
is
pretty
in
the
Python
code.
Anyway,
problems
with
signal
versus
absolute
values
come
to
mind
yeah,
so
absolute
value
is
being
sort
of
a
changing,
a
signal
to
only
output,
a
certain
range
of
it,
but
I
don't
know
yeah.
A
So
this
reads
to
be
a
problem
with
comparisons
and
other
common
and
related
problem
is
shifting
frames
of
reference
during
calculations
picking
the
wrong
intermediate
value
for
further
calculations.
This
could
be
related
to
what
I'm,
trying
to
say
might
be
I
I
feel
is
this
most
suspicious
thing
is
the
fact
that
it
doesn't
know
that
it's
one
period
over
and
over
and
over
again
it
just
it
sees
it
as
a
continuously
building
sequence.
That
will
never
end
the
other.
Two
perennial
problems
are
getting
scaling
wrong.
A
A
Yeah
I
think
you're,
like
probably
bolted
down
to
the
core
problem,
because
every
time
step
in
the
past
is
dependent
on
all
of
the
time
steps
in
the
past
previous
to
it,
and
because
of
we're
seeing
some
randomness
in
this,
because
you
have
to
establish
the
initial
connections
that
are
in
a
random
way,
but
but
I
think
you're
just
seeing
sort
of
a
oscillation
of
the
system
as
it
attempts
to
learn
and
unlearn
about
pattern.
Unlearn
Obul,
because
it
does
not
recognize
that
the
period
is
ending
and
repeating
that's
something.
A
That's
a
deficiency
in
temporal
memory
as
it
is
implemented
today
in
HTM.
U
you
have
to
manually,
tell
it.
This
sequence
has
ended
and
that
sort
of
cut
severs
the
connections
and
then
everything
bursts
again
on
the
next
input.
So
you
can
start
new
sequences
that
make
if
you
were,
if
you
were
doing
that
at
every
step
of
the
way
here.
Reset
reset
reset,
it
would
not
probably
be
doing,
would
probably
get
flat
and
really
confident.
A
A
A
Confidence
values
are
the
ones
that
are
abnormally
fluctuating,
I.
Think
I'm,
pretty
sure
we're
talking
about
anomaly
scores
right
when
you
say
confidence,
values.
Is
that
that's
not
a
term
we
use
I,
don't
know
where
that
came
up.
I
think
Phil
introduced
it,
but
let's
call
it
anomaly
scores
and
that's
what
he's
saying
you
sure
that
always
scores
in
your
plot.
A
What
and
you're
say
it's
either
one
or
zero
okay,
I'm
just
getting
my
term
straight
here,
so
he
added
noise.
The
way
I
suggested
and
it
didn't
really
change
things
very
much
I
mean
there's,
there's
still
it's
not
it's
not!
Well,
actually,
it
does
look
quite
a
bit
more
random.
Doesn't
it
but
I
still
don't
know
what
that
means.
Mark
I'm,
assuming
that
sine
wave
is
the
value
between
1
and
negative
1,
just
for
diagnosis,
diagnostics
or
whatever
try
between
0
&
1
makes
no
material
difference
to
the
results
presents
it.
A
A
A
If
the
encoder
I
don't
know
what
the
encoder
is
for
the
sine,
but
if
it's
a
scalar
encoder,
you
know
if
you
have
to
update
the
min
and
Max
and
stuff
or
it
will
cut
the
input,
you
know,
it'll
only
give
half
of
it:
I
don't
it'll
floor
it
or
ceiling.
It
so
make
sure
that
you've
updated
your
a
coder.
If
you
did
this
for
the
min
and
Max
zero
to
one
and
not
zero
negative
one
to
one
but
I
assume,
that's
where
you're
doing.
A
A
A
A
A
A
A
A
A
A
People
on
Twitch
do
other
people
that
are
like
evangelists
for
Microsoft
or
whatever
they'll
they'll
go
to
a
conference
and
they'll
line
up
some
guests
and
they'll
and
they'll
like
rent
a
little
room
in
the
hotel
and
they'll
have
quick
interviews
with
people
like
in
the
industry.
So
I
was
thinking
about
going
to
some
of
these
AI
conferences
for
a
couple
different
reasons,
one
because
we
never
go
to
these
spaces
I'm
not
going
there
to
learn
about
machine
learning.
A
Would
anyone
be
interested
in
that
because
I
think
that
might
be
fun,
but
I
definitely
don't
want
to
don't
want
to
like
waste
time
prepping
and
going
to
a
conference
and
doing
something
like
this
and
putting
something
live
on
Twitch.
If
it's
like
hey,
we
just
want
to
see
you
in
front
of
your
computer,
okay
cool.
So
at
least
that's
one,
that's
one!
That's
that's
good
enough!
Another
one
is
the
O'reilly
a
I
conference
in
San
Jose
and
that's
I,
think
in
September
they
don't
have
a
too
much
of
an.
They
do
have
some
agenda.
A
Facebook
Berkeley
I
thought
I
knew
that
guy
but
I
didn't
anyway
it
would
be
cool
if
I
can
get
other
people.
If
you,
if
you
look
through,
if
you
recognize
any
of
these
people
on
here
and
you
think
oh
they're
doing
something
related
to
Numenta
they
might,
they
might
be
interested
in
talking
to
us
at
a
conference.
A
Let
me
know
and
I'll
hit
them
up
and
I'll
say:
hey
I've
got
a
twitch
channel
I
talk
about
AI
and
neuroscience,
and
do
you
want
to
come
talk
about
AI
neuroscience
on
the
twitch
show
and
life,
and
so
I
will
do
that.
So
let
me
pop
this
into
chat.
See
here
are
the
O'reilly
speakers
and
the
other
one
was
called:
rework
and
rework
is
like
a
I'd
call:
it
half
half
executive,
half
tech.
So
it's
one
of
those
conferences.
It's
not
a
full-on
tech
conference.
A
It's
got
like
a
big
half
of
its
a
management
track,
an
executive
track
and
they're
talking
about
applying
AI
and
the
strategies
behind
building
products
with
AI
had
all
that
stuff.
So
if
I
did
this
I'd
probably
go
with
our
VP
of
Marketing
Kristy
Naver
and
she
would
like
a
ten
those
tracks
and
I'd
go
to
the
tech
tracks.
We
would
get
a
feel
for
what's
going
on
in
the
industry.
So
that's
the
that's
the
idea
there.