►
From YouTube: Jeff on Minicolumns and SP Part 2
Description
In the last live-stream Jeff did at https://www.youtube.com/watch?v=Z7ot3eNz3aw, I stopped the stream and the conversation kept going! I should have kept rolling, but here is a backup recording from Marcus's computer in case you would like to watch.
A
C
A
A
This
is
what
I
think
is
going
on.
Is
my
interpretation.
There
is
the
edges.
Do
damage
will
see
that
those
accents
are
I?
Can
you
can
fruitfully
find
patterns
there
and
not
get
inhibited
by
somebody
else
so
that
that
was
neuron
just
say,
I'm
gonna
start
representing
those
in,
but
remember
these
actions
are
always
one
thing.
I
didn't
mention
is
accented
damage
they're,
always
like
probing
always
going
in
read
rapidly.
You
know
in
order
days
and
hours,
so
it's
like
little
like
little
worms
they're
going
in
and
out
all
the
time.
A
Represented
by
our
system
will
do
this
if
I
connect
to
them
I
can
get
that
pattern
by
myself
and
I.
Get
depend,
might
be
something
else
going
to
inhibit
me,
so
they
will
actually
want
to
grab
that
information,
and
so
therefore,
what
it
look
like.
The
actions
are
now
representing
moving
out
like
this,
and
then
those
people
start
representing
those
patterns
and
other
people
start
moving
over
said
you
know
everybody
does
or
the
other
way
you
could
just
lose
the
input.
A
So
all
the
neural
tissue
is
sustained,
but
now
you
have
no
input
so
now
you
now
you
have
neurons
that
are
not
getting
any
input
and
they
start
getting
greedy
and
taking
their
neighbors.
So
one
is
one:
is
you
neurons
take
green,
you
start
taking
the
inputs
from
their
neighbors
and
the
other
is
that
under
dead
and
they're
surrounding
their
eye
and
see?
Oh
there's
some
problems.
What
I
can
get
and
they
start
representing
the
whole
I
think
both
of
those
occur
because
you
can
be
covered
from
both
types
of
damage.
A
A
D
A
A
They
get
excitatory
and
inhibitory
inputs,
just
like
pyramidal
cells,
and
they
generate
a
spike.
You
know
just
like
printer
film,
although
their
physiology
can
be
quite
different,
so
they
might
generate
the
first
of
all.
They
can
be
very
fast
to
responding.
Some
of
them,
like
the
basket
cells,
are
extremely.
B
E
A
Varies
quite
a
bit
and
how
they
respond
and,
of
course,
what
kind
of
inputs
they're
getting
so
the
dip.
I'll
answer.
Your
question
is
no
one
really
knows,
but
the
answer
your
question
there's
no,
like
you
can't
say:
oh
the
cells
get
a
fire.
Something
happens.
We've
proposed
some
answers
to
that.
I
was
just
talking
about
how
these
bipolar
cells,
these
double
K
cells,
could
be
recognizing
if
they
recognize
the
same
input
pattern.
Is
they
they're?
Gonna
learn
something,
but
we
really
believe
is
going
on.
A
A
So
these
bipolar
cells
here
connect
generally,
if
I
was
if
I
was
a
pyramidal
Sal
bipolar
cells
make
synapses
on
the
cell
body
and
the
proximal
dendrites,
so
they
have
a
very
strong
effect
on
whether
the
cell
generates
a
spike
it
along.
There
are,
and
so
that's
basic
I'm
going
to
shut
down
this
cell
from
firing
at
all,
there's
another
type
of
cell,
so
that
would
be
basket
cells.
A
This
basket
cells
an
is
bipolar
cells.
Oh
no
I'm
wrong
about
that.
Yeah
one
insect
is
back
and
I'm
confused
here,
yeah
I
think
it's
true
anyway,
there's
another
type
of
cells
called
the
chandelier
cell,
which
we
have
no
idea
what
it's
for
and
it
they
call
the
chandelier
cell,
because
what
it
does
it
makes
connection
along
the
beginning
of
the
axon
yeah,
and
it
looks
like
a
little
candles
candles
and
you
may
say:
well.
Why
would
why
would
a
chandelier
sell?
A
Do
this
and
a
basket,
sell
or
bipolar
scholar
do
this
well
basket
of
the
bipolar
cells
would
prevent
this
cells
from
firing
at
all,
but
would
never
generate
a
spike,
but
it's
possible
and
I'm
just
making
this
up.
It's
totally
of
them.
It's
possible
the
chandelier
cell.
If
it
was
anywhere
these
synapses
actually
formed
here,
it
might
be
possible
for
the
cell
degenerative
spike,
but
not
propagate
down
the
axon
excellent
yeah,
it's
possible,
so
I'm,
just
pointing
out
there
with
these.
These
are
differences.
There's
all
these
very
complex
differences
that
going
on
here
so.
A
A
I'm
not
aware
of
any
I'm,
not
aware
of
any
new,
but
your
neurons
that
are
far
down
in
the
branch.
So
the
the
branch
of
a
cell
is
an
axon.
It's
really
confused
because
it's
gonna
make
five
ten
thousand
synapses
right.
So
there's
a
lot
of
here
I'm,
not
aware
of
any
inhibitory
synapses
anywhere
along
here,
it's
sort
of
like
once
it
gets
past
here
it
goes
everywhere.
There's
no
selectivity
here
that
is
different.
There's
another
class
of
inhibitory
cells.
Oh.
A
Gosh
I,
don't
remember
which
one
is
now
is
it
so
I
talked
about.
I
talked
about
two
times
here
somewhere
right
right
in
the
beginning
of
the
axon.
Some
are
right
near
the
cell
body.
Right
there
are
other
types
of
different
are
cells
which,
if
you
looked
if
I,
if
I,
can
we
just
draw
some
of
these
dendrites
here
right
that
they
are
spaced
along
here
intermixed
with
the
excitatory
cells
that.
A
Yeah,
sorry,
man,
and
so
with
these
you
know
knowing
those
are.
These
are
going
one
one
possible
way
of
thinking
about.
This
is
the
way
I've
always
thought
about.
It
is
well.
Why
would
you
have
these
inhibitory
synapses
along
and
snaps
alignment
and
drives
like
that?
This
would
be
a
way
of
regulating
the
threshold
for
the
dendritic
spikes
right.
So
we
generate
a
generative
spike
and
what
you
could
do
is
you
could
say:
okay,
if
I'm
going
to
make
a
prediction,
the
predictions
can
be
recognized
as
a
pattern
of
activity
out
here.
A
I
can
change
how
easy
it
is.
I
make
a
Bayesian
how,
unless
I
make
information,
so
the
example
I
always
gave
is
I
said
well.
If
you're
looking
at
a
cloud
and
I
said-
oh
sure,
you
see
the
dog
in
the
cloud
of
course,
there's
no
target.
If
I'm
right,
there's
no,
but
you
can
somehow
do
you
focus
your
brain
a
bit
and
also
you
can
see
a
talk.
A
E
A
A
Generalization,
but
this
is
generally,
this
will
allow
you
to
regulate
these,
allow
you
to
regulate
how
many
dendritic
spikes
are
being
generated,
and
so
what
you
want
to
do
you'd
want
to
say
if
I
have
a
very
clear
pattern
or
something
I've
learned
before
I,
don't
need
many
synapses,
but
if
it's
a
fuzzy
pattern
right
now
we're
to
the
cloud
or
it's
different
that
I
saw
before
I
can
just
say:
hey,
you
know
what
that's
all
trying
to
figure
out.
Let's
do
the
best
job
we
can
and
the
best
job.
E
A
The
beginning
of
the
axon
and
then
just
trim
it
around
the
10th
rise
and
we've
never
really
positive
theory
for
this
one,
but
I
have
three
for
that
one.
We
know
we
have
theories
for
this
one,
but
I
always
thought
like.
Maybe
we
could
come
up
with
a
scenario
where
our
selves
need
to
fire.
They
need
to
learn,
but
we
really
don't
want
to
do
anything
well.
I
guess
it
would
be
kind
of
like
imagination.
A
All
right,
I
want
to
think
about
doing
something.
I
want
to
think
about
going
over
there,
picking
up,
match
robot
and
carrying
here.
Well,
I'm,
not
doing
it
but
I'm
thinking,
maybe
I,
don't
know.
Maybe
what
I'm
doing
there
is
I'm,
but
some
reason
you
can
have
cells
it
could
fire
and
learn,
but
not
do
anything.
D
A
A
I,
don't
think
here's
what
I
do
things
going
on
the
basket
cells
with
the
ones
that
we
think
are
doing
it
because
they're
really
really
fast
or
like
super
fast
inhibitors.
They
don't
just
connect
to
a
mini
con.
They
connect
to
multiple
they
connect
to
the
cloud,
and
so
a
winning
cell
would
inhibit
not
just
its
local
salesmen
and
many
column,
but
other
cells
and
other
many
columns,
and
that's
not
because.
A
We
don't
model
it
that
way,
but
but
in
general
you
wouldn't
want
in
the
brain.
You
wouldn't
want
this
mini
column
becoming
active,
and
this
mini
compact
over
Friday.
So
then
either
the
K
cells
help
to
inhibit
each
other,
or
maybe
it's
really.
Maybe
the
whole
thing
is
happening.
Is
that
each
of
the
double
K
cells
but
you're,
the
final
mini
column?
They
might
compete
through
the
basket
cells,
so
they
all
the
best
in
cell
would
come
back
from
then
hit
everyone
else.
Something
like
that.
So.
D
A
Not
it's
not
simply
laid
out
of
the
way
we
did
it.
We
have
a
simplification
of
this
and
that's
almost
all
because
we
decided
a
long
time
ago
we
weren't
going
to
be
modeling
topology,
that
we
weren't
going
to
be
modeling
I,
give
it
this
mini
column,
and
this
when
you
come,
and
this
mini
column
representing
the
same
feature
at
some
spatial
displacement
on
the
resep
on
the
sensory
organ
are
our
don't
have.
A
B
A
Yeah
Yeah
right
and
we
did
that
initially
we
started
off
that
way
and
then
we
decided
that
was
wrong.
Mariana
thinks
that
this
is
really
costly.
Why
are
we
doing
this
cuz?
The
things
we
were
encoding
didn't
really
require
it,
and
and
so
yeah
we
abandon
that
and
then
we
never
went
back
to
it,
but
it
worked
and
it
worked
both
ways.
So.
F
A
The
condition
when
they're
doing
that's
I'm
a
big
blender
or
mouth
yeah,
but
when
they're
that
that
condition
I'm
going
to
point
out
is
the
only
time
I
know
that
occurs
is
when
they're
actually
trying
to
give
like
figure
out
what
these
mini
columns
represent,
like
they're,
they're
doing
those
sinusoidal,
gratings
and
I.
Don't
know
if
that's
true
I,
don't
know
that's
true,
but.
F
A
That's
a
good
point,
there's
other
strange
stuff
going
on
you're
too.
We
now
believe,
of
course,
that
there
are
grid
cells
down
here
and
grid
cells,
as
far
as
we
know
always
have
this
sort
of
bump
of
activation.
You
know
peak
decline,
always,
and
so
some
of
these
cells
down
here
are
representing
a
location
in
a
bumpy
way,
and
that
may
be
actually
what's
driving
the
bumpiness
up
here.
We
don't
really
know
so.
The
there's
a
lot
of
weird
stuff
going
on,
but
from.
F
F
A
I
mean
the
mountain
castle
basically
says:
look
it
shouldn't
matter
what
your
inputs
represent
at
all
right,
there's
gonna,
be
some
within
some
patch
of
sensory
input,
there's
going
to
be
a
spectrum
of
different
properties
and
we're
going
to
do
the
spatial
pull
a
trick
on
all
of
them,
so
they're
all
awkward
innovation
to
divide
up.
Let's
come
up
with
a
set
of
many
columns.
That
is
a
good
of
a
projection
of
that
set
of
properties.
A
Luminosity
or
nothing
right,
I
or
whatever
you
want
to
give
it
and
I'll
just
sort
of
create
an
efficient
map
of
them
all
so,
though,
equally
represented
an
intersection,
so
remember,
I
might
have
several
hundred
mini
columns
within
a
single
column
and
each
mini
column.
Even
though
there's
I
should
have
drawn
this
too
I'm.
Sorry.
A
So
if
you
look
at
actually
I'm
drawn
this
picture
quite
a
few
times,
Marcus
will
know
this.
This
is
the
this
is
the
cube
of
cortex
right
and
so
we've
been
saying:
he's
got
these
little
mini
columns.
Here's
I,
like
you,
know,
Maxwell
skinny.
These
guys
are
you
know
these
are
like
thirty
or
forty
to
one
aspect
ratio
that
really
like
those
things.
A
A
F
A
Or
there's
actually
more
than
two
dimensions,
you
I
mean
it's.
It
seems
like
there's
a
based
division,
which
is
this
guy
and
then
the
other
dimensions
you've
seen
these
maybe
you've
seen
these
a
little
weird
blobby
lava
lamp
like
patterns
that
sort
of
cut
across
all
these
four,
so
you
know
they'll,
be
like
you
know
some
other
pattern.
That's
you
know,
that's
that's
repeating
it
across
some
Ziglar
lines
like
this.
A
The
point
is
you
can
map
multiple
dimensions
onto
this
thing
so,
even
though
there's
a
slab
of
orientation,
individual
mini
columns
in
that
that
all
there's
multiple
mini
columns
of
that
orientation
that
have
different
subsets
or
different
sets
of
the
other
properties,
and
so
when
you
actually
form
a
sparse
activation,
you'd
be
performing
not
just
between
these
mini
comic
book,
between
mini
columns
of
different
things.
We
have
been
doing
with
that.
A
F
A
They
basically
go
by
all
the
cells
that
are
silent
or
that
are
emitting
spikes
that
don't
correlate
to
this
or
making
one
spike
every
once
in
a
while,
and
the
fruit
of
this
talk
to
this
woman
gave
it
cosine
a
couple
years
ago,
which
I
thought
was
very
impressive.
She
made
this
point.
She
said,
like
you
know,
when
we
do
these
experiments,
we're
skipping.
You
know
60
percent
of
the
cells
or
70
percent
themselves,
because
they
don't
and
she
said
that
they
do
a
big
spikes.
Just
not
very
they
don't
seem
to
very
qualitative.
D
A
Or
they
fire
when
we
don't
expect
them
to
fire,
so
she
did
some
sort
of
analysis.
We
she
looked
at
all
those
other
cells
that
she
could
record
from.
That
didn't
seem
to
do
the
right
thing
and
she
showed
that
they
correlated
their
firing,
did
correlate
with
high
or
high
levels
of
experiences,
so
they
weren't
uncorrelated.
They
just
didn't
do
this
basic
firing.
You
know
this
basic,
simple
thing,
which
makes
sense,
I
mean
and
like
if
we
think
some
of
these
are
the
spice
Massell
somebody's
they
could
sell.
A
Some
of
these,
are
you
know,
motor
driven
cells,
whatever
there's
temple
pulling
all
these
things,
they
just
don't
make
sense.
In
this
context,
we
would
expect
them
to
have
different
firing
pattern,
but
you
could
still
have
this
basic
I
did
that
and
now
it's
been
proven
that
there
are
cells
about
every
layer
that
do
have
this
basic
property.
It's
just
it's
just
it's
a
subset
of
all
these
cells.
A
So
there's
some
sort
of
you
know,
there's
some
sort
of
receptive
field
property
that
all
these
guys
and
some
of
these
things
share
in
some
sense,
but
I
do,
but
they
could
be.
They
could
be
representing,
you
know,
location
they
can
represent
their
orientation,
they
can
representing
simple
pooling,
know
all
these
different
things
that
they
could
be
represent.
A
You
know
that
the
column
can
be
representing,
even
though
a
subset
or
being
there's
some
sort
of
unifying
tying
together
of
them
in
some
way,
but
they
actually
doing
lots
of
different
things,
and
that's
we
never
really
and
that's
part
of
what
I
really
want
to
get
to
the
bottom
of.
On
a
theory
point
of
view,
it
is
really
kind
of
a
full
model.
How
do
you
explain
this
structure,
the
different
functions
and
all
the
things
that
the
cortex
has
to
do
and
I
actually
never
closed.
A
A
We
don't
the
way,
I
view
this.
Is
we
don't
learn
to
walk?
We
have
an
undeveloped
nervous
system
and
were
born
were
born
too
early,
so
walking
is
not
controlled
by
the
neocortex.
Walking
is
controlled
by
the
brain
stem
of
the
spinal
cord,
mostly
the
spinal
cord,
and
so,
if
you
were
born,
will
fully
developed
brain
stem
and
spinal
cord.
You
too
would
walk
at
birth,
but
we're
born
premature
now,
I
think
envision.
There's
there's
vision
also
occurs
in
all
parts
of
the
brain,
so
it's
not
just
in
the
neocortex
there's
the
superior
colliculus.
A
That
is
an
old
part
of
that
basically
old,
evolutionary
old
visual
system,
and
it's
been
shown
that
in
primates,
like
our
fear
of
snakes
and
spiders,
is
genetic,
so
you're
born
with
that
and
it's
the
old
part
of
the
brain.
You
don't
to
learn
that
and
it's
the
old
part
of
the
brain.
That
knows
that.
So
the
fact
that
you
know
I
think
some
animals
are
born
and
able
to
start
running
right
away,
doesn't
mean
they
know
the
world.
A
D
A
I,
don't
think
those
are
samples
are
just
going
to
counter
to
the
whole
theory
here
they
just
you
really
have
to
just
a
win
with
thee.
What
was
the
maturity
of
the
old
of
the
part
of
the
brain
and
the
approaches
still
hasn't
learned
anything
the
coward
will
probably
cannot
recognize.
You
know
companies
certainly
not
gonna
how
to
get
to
the
waterhole.
C
A
C
A
Because,
well
the
one
of
the
observations
that
some
of
these
basic
orientations
are
there
at
birth
and
so
like.
Well,
how
could
they
be
learned
right
there
at
birth
and
the
hypothesis
is
well
is
that
in
in
when
the
child
is
developing
in
utero,
there
are
these
waves,
as
you
said,
although
that
that
lines
of
activation
that
travel
across
the
retina
at
different
orientations-
and
they
wouldn't
theory-
could
produce
this
learning
here.
A
But
you
know
that
again
so
then
tell
people
gone
to
the
extents
that
oh
well,
that
proves
that
these
aren't
learned,
which
is
it's
whole
because
you
could
we
learn
these,
so
they
you
know
so,
just
because
you're
born
with
some
basic
abilities
at
Birth,
but
you
know
I,
don't
think
there's
any!
No
one
I,
don't
think
anyone
believes
that
you're
boring
at
birth
being
able
to
recognize
the
chair.
A
I
mean
who's
like
maybe
you
can
start
with
a
little
bit
of
lines,
but
if
so
again
that
could
be
an
evolutionary
advantage
just
to
save
the
animal
from
having
this
bed
I'm
out
of
time
learning
these
basic
functions.
They
could
be
born
with
it,
but
he
has
taught
the
learner.
So
it's
not
really
a
counterexample.
That's
like
a
weird
one,
but
some
people
are
all
these
can't
be
there
because
they're
you
I've,
had
people
tell
me
they're
a
scientist
to
my
face
like
I,
don't
believe
he
wanted
learning
anything.
A
It's
the
person
who
studied
to
be
one
I'm
like
you
got
to
be
joking.
It's
a
lawyer!
You
just
reach
into
the
cortex.
You
don't
think
it's
learning
anything
and
one
of
the
arguments
it's
like
well
when
you're
born
or
two
noses.
So
what
do
you
put
so
gonna
learn?
You
know
that's
a
very
simple
idea,
but
I
think
they're
talking
on
because
you
can't
you
can't
beat
none
of
these
things
to
trauma
or
different
patterns
of
employers
could
innovate
it
with
something
else.