►
From YouTube: 21 March 2023 Overview of Neuromorphic Computing Challenges and Opportunities by Sakib Hasan PhD
Description
Neuromorphic Computing promises orders of magnitude improvement in energy efficiency compared to the traditional von Neumann computing paradigm. The goal is to develop an adaptive, fault-tolerant, low-footprint, fast, low-energy intelligent system by learning and emulating brain functionality which can be realized through innovation in different abstraction layers including material, device, circuit, architecture, and algorithm. As the energy consumption in complex machine learning tasks keeps increasing exponentially due to larger data sets and resource-constrained edge devices becoming increasingly ubiquitous, neuromorphic computing approaches can be a viable alternative to a deep convolutional neural network that is dominating the field today. In this talk, I introduce neuromorphic computing, outline a few representative examples from different layers of the design stack (devices, circuits, and algorithms) and conclude with a few important challenges and opportunities in this field.
C
B
B
B
he
received
his
PhD
degree
in
electrical
engineering
from
the
University
of
Tennessee
Knoxville
in
2017..
Dr
Hassan
is
doing
active
research
on
neuromorphic
Computing,
secure,
Nano,
electric
circuit
design,
semiconductor
device,
modeling
and
non-linear
Dynamics
and
vlsi
circuit
design.
In
today's
talk
he
will
introduce
us
with
the
basics
of
neuromorphic
computing
and
also
talk
about
his
own
research.
Additionally,
he'll
give
us
some
idea
about
the
important
challenges
and
opportunities
in
this
field,
so
without
further
Ado
Dr
Hassan.
The
floor
is
yours.
D
Thank
you
so
much,
first
of
all,
I'd
like
to
thank
you
for
giving
me
this
wonderful
opportunity
and
arranging
these
things
by
one
week.
So
today's
talk,
of
course,
I
wanted
to
first
introduce
other
audience
to
network
with
Computing.
It
is
a
very
hot
research
topic
right
now.
There
is
a
lot
of
interest.
It
has
any
like
any
Computing
Paradigm.
It
has
a
lot
of
different
abstraction
layer,
so
covering
everything
in
details
is
beyond
the
scope
of
a
one
hour
clock.
D
So
I
will
try
to
give
a
summary
talk,
but
I
am
going
to
commit
this
from
my
background,
so
it
might
be
different
from
someone
who
is
coming
from
a
more
computer
science
and
algorithmic
background.
So,
given
that,
so
let
me
start
with
this
thing,
we
have
a
lot
of
slides
so
since
I
have
a
lot
of
slides
and
two
slides
can
contain
a
lot
of
you.
A
D
Information
one
thing
I
would
first
like
to
make
clear:
is
that
what
I
expect
my
audience
is
that
I
am
not
assuming
anything
so
I'm,
assuming
that
it
would
be
a
wide
ranging
audience
so
sometimes
I
can
go
over,
something
which
seems
very
basic,
and
sometimes
it
feels
like
I'm
going
over
things
too
fast,
so
reason,
I'm,
organizing
the
talk.
This
way
is
that
Dr
mahabu
told
me
that
this
will
be
recorded
and
these
will
be
uploaded.
D
So
the
idea
of
the
slides
is
I'm
not
expecting
you
to
follow
the
slides,
all
the
slides
and
all
the
information
you
don't
actually
need
to.
It
might
be
overwhelming
YouTube
better.
Can
you
can
take
it
like
a
podcast?
So
if
you
just
follow
what
I'm
saying
and
kind
of
follow
the
story,
I
think,
then
you
will
get
the
big
picture
and
that's
what
I
am
getting
now
the
reason
I'm
putting
all
this
information
in
the
slides
is
that
it
will
be
recorded
and
I'll
also
provide
the
slides
with
all
the
resources.
D
D
Okay.
So
with
that,
let
me
start
with
the
topic,
so
first
I
am
going
to
talk
about
before,
starting
with
neuromorphic
competing
I'll
talk
about
Computing
in
general,
so
what
kind
of
computing
we
are
currently
doing?
So
you
guys,
probably
most
of
you-
are
familiar
with
this-
that
Computing
is
a
huge
thing.
D
It
has
changed
the
world
in
the
last
70
18
years,
especially
with
the
great
pioneers
like
porn
women
and
doing
the
idea
of
computation
became
more
rigorous
with
shown
on
the
idea
of
information
people
more
regular,
and
then
we
have
this
wonderful
jackpot.
We
get
the
jackpot.
We
came
up
with
the
transistor
and
then
start
things
getting
started
miniaturized
and
now
we
have
almost
unbelievable
computing
power
at
our
hands.
Sometimes
maybe
we
take
it
for
granted.
This
is
amazing.
D
The
way
this
field
has
evolved,
I
would
be
as
daring
to
say
that
it
has
never
happened
in
the
history
of
human
with
any
other
technology
and
I'm,
not
saying
that,
because
it's
my
field,
I,
think
this
is
objectively
true
and
not
because
electrical
and
computer
Engineers
or
computer
scientists
are
more
smarter
than
other
fields
of
engineering.
I,
think
we
we
got
lucky.
We
really
got
lucky
with
this
some
of
these
things,
but
of
course
there
are
a
great
deal
of
Engineers
went
into
this.
D
So
I
first
want
the
audience
to
appreciate
the
what
happened
in
the
Computing
Revolution,
so
any
Computing
system,
the
kind
of
computer
system
that
we
use
nowadays,
mostly
digital,
mostly
blsi,
born
human-based
system.
It's
not
possible
to
work
with
billions
of
transistors
that
we
have
in
a
single
chip,
so
we
use
abstraction
layers,
and
this
is
the
kind
of
figure
people
used
so
at
the
very
low
level.
If
you
follow
my.
D
D
So
of
most
popular
is
silicon,
but
of
course
you
can
use
the
other
materials,
so
people
have
to
some
people
work
on
this
field,
they'll,
develop
the
materials
and
the
device
and
characterize
it.
On
top
of
that,
there
is
a
circuit,
so
you
combine
some
of
this
device
to
build.
Useful
certificate
can
be
digital
circuit.
It
can
be
analog
circuit
in
the
digital
domain.
D
D
Modeling
dnsi,
that
kind
of
thing
and
in
2017
I
went
into
a
postdoc
in
the
University
of
Tennessee
After
Effects
PhD
under
Dr
Garrett
Rose
in
10
I
mean
utk,
and
that's
why
I
got
used
to
got
like
introduced
to
this
neuromorphic
Computing
thing
which
I
found
fascinating
so
but
before
going
there,
I
need
to
show
like
why
we
need
University
company.
So
this
was
great.
This
conventional
Computing
Paradigm
was
working
great,
so
it
all
builds
upon
plsi.
You
can
build
a
lot
of
transistors
into
this
thing.
D
I
am
not
going
to
bore
you
with
all
the
details,
but
surprise
it
to
say
that
a
single
transistor
you
can
turn
it
on
or
off.
It
can
be
one
or
zero
logic.
Based
on
that,
you
can,
if
you
can
put
a
lot
of
them
together
and
make
them
very
tiny,
then
you
have
a
lot
of
computer
in
your
hand.
So
there
was
the
base
basic
idea.
D
The
growth
rate
is
phenomenal.
As
I
have
mentioned,
this
is
kind
of
a
plot
of
universal.
As
you
can
see,
it's
going
on
has
been
going
on
for
a
long
time
now
now,
a
little
bit
of
background
in
the
invention
of
the
transistor,
which
I
think
is
a
game.
Changer
is
1947
in
Bell
Labs,
three
scientists
first
came
up
with
a
transistor
before
we
used
decorum
to
you
before,
then
we
used
even
bigger
switches.
D
Transistor
is
the
very
small
thing
it
can
work
as
the
switch
and
if
you
can
put
a
lot
of
them
together,
you
can
build
useful
circuits.
So
initially
people
cried
with
something
called
bjts
bipolar.
Transistors
then
came
mosfet.
There
are
a
lot
of
challenges
with
mosfet
people
figured
it
out.
We
had
nmos
Emos,
and
then
people
come
up
with
the
idea
of
complementary
mosque.
Cmos
in
1960s
turned
out
that
for
digital
circuit,
it's
awesome
because
no
aesthetic
power
loss,
so
it
only
computes
Power
when
it
is
doing
any
real
competitions.
D
Input
is
changing
only
that
Dynamic
problem
no
static
power.
Some
of
these
assumptions
are
not
true
anymore.
In
the
last
15
years,
things
have
been
drinking
down
or
transistors
have
become
more
and
more
leaky,
which
will
actually
lead
why
we
need
a
new
Computing
pattern.
So
that's
this
is
kind
of
the
pattern.
Okay,
so
this
integrated
circuit
1958
people
learned
how
to
put
a
lot
of
this
transistor
together.
It
was
a
big
thing,
another
Nobel
Prize
on
these.
D
We
have
now
all
these
internal
and
Indies,
and
this
huge
microprocessors
with
billions
of
transistors
together
in
integrative
circuit.
We
are
kind
of
used
to
IC
chips
nowadays
in
every
set
of
online
now
I
would
mention
something
which
probably
all
of
you
have
hired,
is
called
more
slow.
The
more
slow
is
not
really
a
law,
it's
just
an
observation,
Intel
founder
who
made
in
1965
that
he
saw
that
he
only
had
four
data
points.
D
Actually,
as
you
can
see
at
that
time,
and
he
saw
that
the
number
of
transistors
integrated
circuit
has
been
kind
of
doubling
every
18
months,
so
he
predicted
that
this
would
probably
go
on
based
on
the
technology
for
the
next
two
decades,
which
was
enormous.
If
it
actually
happens,
nobody
expected
an
exponential
group
to
grow
more
than
20
years
into
20
years
was
a
huge
extra
policy.
D
Of
course,
last
10,
15
years
it
has
been
slowing
down,
which
is
why
we
need
other
company
Paradigm,
but
as
I
just
want
to
emphasize
that
this
has
been
I
still
find
it
mind-bending
that
this
has
actually
gone
on
for
last.
So
just
to
give
an
example,
like
the
big
thing
in
1971
was
the
Intel
microprocessor
commercial
microprocessor,
it
was
a
huge
Thicket.
We
have
200
or
2300
switches,
2
300
switches,
transistors
energy,
so
nothing
people
have
not
seen
nothing
like
this.
D
It
was
a
huge
thing,
2000
cheap.
So
that's
a
three
orders
of
magnitude.
Now
we
have
nine
orders
of
language
or
billions
of
times,
so
we
have
increased
the
number
of
switches
by
six
order
of
magnitude
from
1971
to
now.
So
let's
say
50
years,
which
is
and
I
don't
have
much
word
for
this,
like
very
surprising
I
mean
no
field
has
grown
like
this
anyway.
D
So
sometimes
what
happens
is
that
when
you
have
a
field
which
is
going
so
fast
people
kind
of
get
used
to,
and
nowadays
everybody
expect
that
every
two
years
everything
will
be
double.
You
know
this
move
slow
thing,
so
you'll
have
more
transistors
in
the
cheap,
so
transistor
compute
cost
will
be
less
you'll,
have
more
complete
power,
more
memory
and
everything
and
people
can
never
expect
that
every
two
years
this
will
happen.
D
If
that
does
not
happen,
people
get
impatient,
but
actually
the
fact
that
this
has
been
happening
for
last
56
years.
That
is
the
miracle,
not
the
fact
that
it
has
been
slowing
down
in
the
last
10
15
years.
I
think
has
been
going
great,
so
this
is
showing
for
40
years
friend.
So
it's
a
semi-lock
block,
so
linear
and
semi
lock
means
that
exponential
I
think
all
familiar
with
this
kind
of
thing.
D
D
A
D
So
study
10
micrometers,
so
now
people
are
making
10,
nanometer
I.
Think
tsmc
is
also
making
seven
nanometer.
Global
is
also
I
think
to
it's
difficult
to
keep
track
of
all
these
technology
board,
because
every
two
three
years
they
come
up
with
anything
but
I
think
we
are
roughly
hovering
around
10
nanometer,
seven
nanometer.
The
problem
is
that
this
has
been
very
difficult,
but
because
booster
has
a
positive
feedback
kind
of
thing
so
because
there
is
huge
investment,
people
are
buying
a
lot
of
chips.
So
there
is
huge
investment
materials
device
scientists.
D
They
have
been
working
and
always
finding
a
way.
So
when
people
heat
one
micrometer
around
1980s
people
thought
that
one
micrometer
is
kind
of
the
legal.
How
small
can
you
make
it?
So?
One
micrometer
is
1000
nanometer
right
and
we
have
now
a
10
elements,
so
one
micrometer
one
seemed
impossible.
We
went
below
100
nanometer,
which
also
seemed
impossible,
and
people
predicted
this.
D
So
there
is
a
funny
saying
that
the
people,
the
number
of
people
that
say
Rose
law
is
there
goes
exponentially
every
year,
because
this
thing
they
are
predicting
that
Moore's
Law
is
ending
more
slow.
Is
ending,
it
has
been
going
on
for
the
last
30
years.
People
are
saying:
no,
it's
not
possible.
You
cannot
go
below
100
nanometer,
you
don't
have
the
word,
then
you
cannot
do
the
photos.
D
Engineers
have
always
found
a
way
around
this
right.
So
and
people
kind
of
I
think
got
complacent
a
little
bit.
They're
always
there
will
be
clearly
some
kind
of
way
around.
However,
things
have
become
really
difficult
under
19
million
microsystem
90
and
especially
under
65
45
nanometer
thing.
First
of
all,
the
transistor,
our
conventional
physics
of
transistor,
the
way
it
works
has
become
like
it
has
become
leaky.
So
the
main
advantage
of
CMOS
was
that
the
static
power
loss
has
almost
seen
it's
not
true
anymore.
D
D
Effect
very
close
to
that
a
lot
of
quantum
effects
become
prominent,
so
tunneling
and
other
non-ideal.
It
is
become
bad
for
you
something.
So
we
have
been
facing
issues
a
lot
of
issues
so
nowadays,
going
from
one
technology
node
to
another
technology.
Nodes
is
a
huge
investment
so
that
fabrication
facilities
that
we
have
have
shrinked
and
free,
and
now
we
only
have
a
handful
of
fabrication
facilities
who
can
make
this
modern
day,
14
nanometer
10
nanometer
technology.
D
So
there
is
this
tsmc,
of
course,
the
big
chant
now
and,
of
course,
interested
makes
their
own
jeeps,
and
that
is
Samsung
and
global
I
think
I
didn't
make
something,
but
very
few
okay.
So
there
is
a
lot
of
security
and
U.S
government
everybody,
so
they
passed
the
chief
sex
recently,
so
I
want
to
make
onshore
chips
and
all
that.
So
there
is
this
huge
problem
with
scaling
down.
I
mean
it
is
still
going
on,
but
it
has
been
slowing
down
and
there
is
a
lot
of
problems.
People
are
facing
I
personally
think
again.
D
I
might
be
proven
wrong,
but
I
personally
think
that
this
time
around,
we
are
actually
in
a
real
bind.
It's
not
like
the
previous
era.
Well,
it
was
a
technological
problem
like
how
can
you
make
below
100
below
five
nanometer,
making
a
good
transistor
using
the
physics
that
we
are
using
now
for
Monster
or
finfet
nowadays,
I,
don't
think
how
it's
possible
I,
maybe
again
is
From
naive's
perspective,
maybe
but
I,
don't
think
I
think
this
time.
This
is
real.
I
think
this
time,
mostly
is
actually
in
and
we
should
prepare
for
it.
D
So
the
implication
of
Ursula
is
that,
even
if
you
don't
make
any
changes
in
the
software
in
the
competition
and
everything
everything
Remains
the
Same
just
because
transistors
are
becoming
half
and
half
size,
you
will
get
double
or
double
apartment.
Of
course,
computer
Architects
have
been
great
works
and
software
engineers
and
algorithmic
work,
but
even
if
you
don't
do
anything
just
because
of
rules,
the
computation
power
has
been
going
down,
you
know
I,
don't
think
we
can
rely
on
this
or
maybe
it
will
go
on
for
next
10
decade,
albeits
slower,
but
I.
D
D
D
A
few
other
slides
which
explain
this
situation,
why
this
is
important
as
a
historical
perspective,
I'm
not
going
to
go
into
any
details
on
these,
but
let
me
go
into
something
of
like
like
why
we
need
neuromorphic
company.
Even
this
background
after
this
particular
so
I
have
to
put
some
more
slides.
So
if
anybody
is
interested
can
I'll
upload
the
slides
as
well
with
the
videos,
so
you
can
look
at
it.
D
But
now,
given
this
background,
that
conventional
computation,
Paradigm
has
been
slowing
down
and
all
that
so
move
slow
has
been
slowing
down,
people
have
been
looking
into
alternative
techniques,
so
is
it
only
the
fact
that
root
slow
is
slowing
down
not
freely?
So
there
are
two
other
factors
that
has
emerged
in
the
last
20
minutes,
the
first
of
them
I
think
the
most
important.
Of
course
I
have
talked
about
the
plateauing
of
moves.
You
know,
but
the
another
important
thing
is
called
Von
Newman
bottle.
D
So
if
you
probably
all
of
your
family
agree
with
that,
the
conventional
computer
digital
computer,
that
you
use
it
uses
an
architecture
originally
developed
by
John
woman
and
his
colleagues.
So
it's
called
the
bomb
demon
architecture.
The
thing
that
we
need
to
know
for
this
job
is
that
in
one
human
architecture
there
is
a
clear
divide
between
memory
and
computation.
There
is
a
processor
which
does
computation.
D
There
is
something
where
you
store
the
information,
so
if
you
as
a
human
being,
you
need
to
remember
things
and
then
you
have
to
make
reasoning
and
make
decisions
in
the
human
brain.
Those
things
are
located
in
the
same
place,
but
in
digital
computer
that
people
develop
in
the
early
40s,
because
architecture
was
clearly
separated,
so
memory
and
processing
were
separate.
Now,
rules
has
been
going
going
up
faster
and
everything,
so
computation
has
been
fasting,
beginning
faster
and
faster
processor
and
memory
board.
D
However,
memory
has
not
been
becoming
as
fast
as
the
computation
that
processor
is
moving,
so
what
what
is
happening
is
actually
counterintuitive.
Most
of
the
time
digital
computer
currently
takes
to
solve
a
problem
is
not
really
doing
the
actual
competition.
It's
going
back
and
forth
between
memory
and
the
process.
This
is
called
the
war
Newman
model,
so
people
have
been
trying
to
come
around
like
how
can
you
solve
this?
How
can
improve
the
memory-
and
there
has
been
great
work
of
this,
but
this
is
a
fundamental
problem
of
the
architecture.
D
Again,
it's
not
about
like
deeply
beneficial.
The
architecture
itself
is
like
this
and
monument
and
the
early
Pioneers.
They
are
always
looking
into
human
brain
and
trying
to
inspire.
They
knew
that
there
is
another
way
of
doing
this,
but
we
started
with
this
thing.
It
is
simple
to
understand
and
with
more
slow
exponentially
doing,
we
have
did
not
have
to
worry
about
this
thing
that
starting
from
2005
this
became
really
prominent
people
starting
becoming
more
notice
in
the
snow.
Another
reason
that
happened
is
2005
2005
2006..
D
We
started
hitting
something
called
a
power,
so
this
is
sometimes
called
the
memory
one
not
even
better.
There
is
another
world
that
we
are
hitting,
which
is
called
the
power
one.
So
this
is
the
third
point
that
I'm
saying
so
power
wall.
Again
the
details
is
a
little
bit
more
subtle,
but
for
this
Talk's
purpose.
So
what
is
happening
is
that
when
you
do
the
scaling,
it's
more
slow
thing,
making
transistors
smaller
and
smaller,
there
is
something
called
Denard
scaling,
so
it's
pioneered
by
dinner.
D
So
it
means
that
yes,
transistor
becomes
smaller,
so
you
can
pack
more
of
them
in
into
the
same
area,
but
also
they
become
faster
because
electron
has
to
travel
less
and
also
the
power
of
this
thing
also
goes
down.
So
the
overall
power
density
Remains,
the
Same,
so
the
chip
does
not
become
too
hot,
so
this
is
again
that's
why
I'm
saying
that
I
think
we
hit
a
jackpot
in
this
technology,
because
usually
it
never
happens.
If
you
make
a
change,
there
is
a
usual
trade-offs.
One
thing
goes
down.
Another
thing
goes
up.
D
A
D
Now
around
2000.
What
problem
we
had
is
that
when
you
go
with
100
nanometer,
it's
very
difficult
to
scale
everything
down
you
can
make
the
transistor
smaller.
That
does
not
mean
you
can
reduce
the
threshold
voltage
in
the
same
scale,
so
threshold
voltage,
you
can
think
of
that
transistor
is
a
switch.
You
have
to
apply
a
certain
voltage
to
turn
it
on
or
below
that
it's
off.
So
that's
that's!
A
thirst
Supply
should
work.
D
What
we
expect
is
that
everything
will
go
down
equally,
it
does
not
happen
so
after
2000,
this
problem
became
really
serious
because
there
are
some
inherent
limit
again.
This
is
not
a
technology.
This
is
the
limit
of
the
fundamental
physics,
our
mosfet
physics,
the
way
it
works.
It
is
restricted
by
a
fundamental
physical
law
which
is
called
the
boltzmann
limit,
so
in
this
particular
physics,
the
sub
threshold.
D
So
when
you
make
the
transistor
off
you
go
below
threshold,
the
minimum
you
can
achieve
60
millivolt
per
decade,
which
means
that
if
you
want
to
reduce
the
current
by
a
decade
which
means
10
times,
you
have
to
reduce
the
voltage
by
60
millimeter,
at
least
no
real
transistor
achieve
this
ideal
thing
is
the
best
one
may
be
1900.
Something
like
that.
D
You
can
do
some
double
gate
thing
to
do
which
of
these,
but
you
know,
let's
say
the
best
possibilities:
60
millivolt
per
ticket,
to
distinguish
a
good
one
and
zero
at
the
noise
immunity,
which
was
the
huge
thing
why
digital
one
over
n?
You
need
to
have
a
good
distinction
between
ones
and
C,
so
you
need
to
have
a
let's
say
five
times
three,
so
60
millivolt
times
5
is
300,
so
you
need
at
least
to
have
this
300
multiple
cap,
and
if
you
stack
more
transistors,
you
need
even
more
power
supply,
long
story
short.
D
What
people
found
out
after
2000
is
that
vdd,
the
power
supply
of
the
Jeep
is
not
scaling
down
the
way.
Transistor
is
stealing
them.
Why
this
is
so
important,
because
is
it
turns
out
that
the
energy
you
all
know
is
half
CV
Square,
the
capacitor
energy,
so
the
supply
voltage
energy
to
do
some
computation
will
be
CV
square
and
then
you
know
you
can
go
into
Power
by
multiplying
with
frequency,
but
the
main
story
is
that
the
energy
is
compared
proportional
to
V
Square,
dot
V.
D
So
if
V
does
not
go
down
in
the
same
Speed,
Energy
actually
does
not
go
down
in
B
squared,
so
energy
becomes,
the
power
becomes
significantly
more
then
it
can
contain
in
that
chain.
So
if
you
actually
keep
the
scale
you
can
put,
let's
say
two
times
more
transistor
power
should
remain
roughly
the
same,
doesn't
appearance
it
doesn't
particularly
keep
increasing.
So
around
2006,
the
Chiefs
started
became
so
hot
that
it
was
not
feasible
anymore.
So
before
that
what
people
used
to
do?
D
You
probably
remember
this
I
remember
this
from
my
Early
Childhood
that
whenever
the
computer
they
came
up
with
a
new
computer,
they
always
promoted
at
the
clock
frequency
before
it
was
one
gigahertz.
Now
it's
two
gigahertz,
that's
open
around
2005
things
started
changing
block
frequency
kind
of
became
saturated
because
if
you
increase
the
clock
frequency
even
further,
the
energy
will
be
too
much
to
control.
You
know:
Power
Hit,
That,
Power
people,
I
mean
Intel,
actually
had
a
very
difficult
time
with
that
cheap
I
forgot
the
name
around
that
time.
D
So
there
was
a
c
change
in
the
computation
world.
At
that
time,
people
found
out
that
this
whole
clock
frequency
increasing
into
single
CPU.
It's
not
sustainable.
You
can't
do
this
anymore,
even
if
mood
slow
continues,
which
is
this
not
but
I
mean
even
is
continuous
because
of
this
power
one.
We
could
not
increase
that,
so
people
came
up
with
a
different
architecture
like
multi,
multi-core
CPUs,
and
from
that
it
has
become
the
norm.
So
two
core,
now
four
core
eight
cores.
D
Now
we
had
huge
number
of
server
cores,
so
parallel
Computing
is
kind
of
a
circumvent
in
this
problem.
Okay,
so
these
are
I
would
say
that
the
three
main
problems
that
facing
the
conventional
computing,
so
people
are
becoming
worried
because
a
lot
of
disease
has
to
do
with
fundamental
physics.
So
how
can
we
get
around
this
now?
One
simple
solution:
is
there
like
what,
if
we
come
up
with
the
new
terms,
a
new
Switch
which
does
not
have
any
of
this
problem
does
not
have
both
limit.
D
Will
solve
all
our
problems
so
the
next
day
some
research
there
comes
up
with
a
new
transistor,
which
is
a
new
physics,
and
people
have
been
promising
this
thing
for
last
15
years.
Most
of
them
have
not
been
realized
yet
so,
for
example,
there
have
been
some
exciting
work
in
negative
capacitance
transistor,
which
theoretically
can
go
beyond
the
full
spending.
So.
E
D
Are
very
exciting
work
on
this
field,
maybe
some
of
them
will
become
fruitful,
we'll
see,
but
so
that's
very
like
device
level.
If
you
come
with
a
new
device,
new
Switch.
Similarly
from
back
home
to
you
to
transistor,
maybe
we
will
go
from
transistor
to
a
new
device
and
that
will
change
everything
possible,
so
some
people
are
exploited.
Other
people
are
going
from
instead
of
going
Bottom
up
going
from
top
down.
That
may
be
the
way
we
are
doing
the
computation.
You
need
to
rethink
our
one
human
architecture.
D
Maybe
there
is
a
different
way
of
doing
computation,
and
this
is
where
they
came
up
with
the
idea
of
neuromorphic
computing
or
brain
inspired
completely.
Of
course,
you
guys
probably
have
heard
about
Quantum,
Computing
and
other
Paradigm.
Those
are
also
very
exciting
things
and
people
are
exposing
them,
but
a
natural
thing
to
ask
is
that
is
the
digital
computer.
The
way
it
is
Computing
is
the
only
way
to
do
competition,
obviously
not
because
we
human
beings
or
other
members.
D
So,
yes,
maybe
we
are
not
as
precise
or
as
fast
as
digital
computers,
but
we
are
very
good
at
certain
tests
and
computers
are
very
good
at
certain.
Digital
computers
are
very
good,
so
maybe
for
some
of
these
tasks,
like
natural
language,
processing
or
image
processing,
where
Evolution
over
millions
of
years
have
prepared
out
this
organisms,
this
brain
to
solve
this
problem
very
efficiently,
very
low
energy
cost.
D
Maybe
we
can
learn
something
from
that
and
build
a
modern
day
computer,
which
is
a
very
different
architect.
It's
not
like
a
non-newman
architecture.
One
memory
and
processing
is
separate.
It's
more
like
brain
organisms,
brain
where
things
are
co-located
in
it
consumes
very
low.
Energy
has
a
very
different
algorithm
and
architecture
of
computation,
and
maybe
then
we
can
come
up
with
new
devices
which
will
enable
this
new
technique.
So
it's
more
from
the
top
down
for
students,
but
I
mean
to
do
this.
I
want
to
emphasize
one
figure
that
I
just
mentioned
before
that.
D
D
And
you
know
application
software,
which
is
not
shortly
same
thing
with
neuromorphic
Computing,
so
it's
not
possible
for
me
to
cover
all
of
them
in
any
detail
and
I'm,
not
qualified,
actually
to
talk
about
all
of
these
things.
I
work
on
a
certain
particular
development
system,
but
I
will
try
to
cover
at
least
everything
to
Peak
your
interest
so
that
you
can
explore
which
staff
field
you
are
into
anyway.
So
this
is
the
main
internet's
brain
inspired
Computing,
but
there
are
some
other
interests,
so
something
very
interested
happened
around
2010
and
I.
D
Think
all
of
us
are
familiar
with
is
that
this
artificial
intelligence
has
been
happening
for
a
long
time.
People
have
been
exploring
for
many
products.
Neural
network
is
also
not
new,
the
main
algorithms,
even
the
back
propagation,
it
is
90s.
People
knew
this.
It
was
not
the
most
popular
one
because
we
did
not
have
enough
data
and
it
did
not
have
enough
computation
brings
back
has
to
move.
Actually.
D
B
D
Network
became
the
range
and
still
is
a
huge
thing-
chair
GPT
nowadays,
everybody's
talking
about
this,
so
so
neural
network
is
of
course,
people
talking
about
again.
Neural
network
is
a
pain
in
spared
computation
and
some
people.
When
people
talk
about
neuromic
committee,
some
people
also
include
conventional
neural
network
like
deep
neural
network
or
combination
of
neural
network.
I
will
back
to
different
a
little
bit,
but
this
is
a
good
application.
D
So
so
there
are
many
people
who
would
actually
include
this
as
part
of
minimum
of
the
Computing
as
well
so
efficient
implementation
of
neural
network,
because
neural
network
is
by
nature
parallel.
So
if
we
want
to
implement
that
in
our
sequential
computer,
it
takes
a
lot
of
energy,
it
is
also
not
very
efficient
and
that's
why
actually,
they
have
to
wait
until
2010
when
gpus
are
available
and
they
have
these
people
who
can
almost
magically
program
this
gpus.
To
do
this
thing,
so
the
conventional
Computing
Paradigm.
D
The
way
we
land
programming
is
a
sequential
thing.
It's
not
really
suitable
for
neural
network
kind
of
thing,
more
brain
like
thing
because
it
is
originally
inspired
from
brain.
So
if
you
could
make
something
inherently
parallel
in
Computing
architecture,
which
is
inherently
parallel
and
where
things
are
memory
and
computation
is
co-located,
maybe
we
can
do
efficient
implementation
of
digital
Network
and
people
have
been
exploring
new
devices.
D
I'll
mention
a
few
of
them
and,
of
course,
the
convergence
of
efficient
offline
and
online
learning
in
this
new
world
of
iot
and
beat
data,
so
in
iot
I'm
mentioning
because
billions
of
devices
are
now
connected
together,
intelligence.
We
want
Intelligence
on
these
devices,
but
we
are
very
power,
limited
in
iot
battery
power
limited.
So
we
don't
want
them
to
run
for
too
long.
So
we
want
to
make
them
as
energy
efficient
as
possible,
as
I
have
mentioning
that
conversational
Computing
consumes
a
lot
of
energy.
D
Our
brain
consumes
a
lot
less
energy,
so
you
probably
have
heard
that
alphago.
So
there
is
this
machine
learning
thing
defeated
the
winner
of
alphago
yeah.
That's
one
way
of
looking
at
it.
Yes,
but
if
you
look
at
how
much
energy
that
world
champion
in
go
is
consuming
while
he's
doing
and
what
is
the
Machinery
algorithm
is
consuming
by
history,
it's
orders
and
Orders
of
magnitude.
D
So
the
question
is
computation,
but
at
what?
What
cost
we
often
times
when
people
in
the
competition
world
were
so
focused
on
accuracy
and
performance
speed
they
ignore
they
used
to
ignore
the
energy
aspect
of
it,
but
at
this
stage
I
don't
think
it
is.
We
can
ignore
this
anymore.
So
that's
another
inspiration,
audio
fee
competing
which
promises
that
are
similar
to
human
brain
or
mammalian
brain.
We
can
call
compute
efficient
with
less
okay,
so
that's
kind
of
the
motivation
behind
neuromorphic
computing.
D
So
there
is
a
big
picture
very
rough
picture
here.
Let
me
go
ahead
so
for
the
resources.
So
in
this
talk,
I
am
like
again
all
the
stakes
in
the
computation
layer.
I
cannot
cover.
So
I
am
following
a
summary
review
people,
so
this
is
from
koshikai
and
priyadoshini
Panda,
so
it's
recently
published
in
nature,
so
they
make
a
summary
perspective
paper
or
neuromorphic
Computing
I
found.
As
going
to
a
few
of
these
review
articles.
I
found
this
one
nicely
covered.
D
What
I
wanted
to
discuss
with
them,
but
mostly
spiking
euromorphic
system
and
so
I'll,
follow
this
article
mostly
and
if
this
has
a
lot
of
great
references
to
die
to
the
seminal
papers
in
the
field.
So
if
you
are
interested,
you
can
always
go
and
read
this
paper
and
check
those
references.
If
you
want
to
go
deep
into,
let's
say
algorithm
or
device
or
circuit,
or
you
know
any.
A
D
D
Now
there
is
actually
a
lot
more
going
on
going
on
in
actual
human
weight,
but
for
our
purposes,
let's
just
keep
it
simple
and
say
that
neuron
is
like
the
main
computational
layer
in
it,
and
synapse
is
the
main
learning
in
actual
brain,
not
as
simple
as
that
scenes
to
a
lot
of
population
as
well.
You
can
also
can
learn
and
end
up
with
time,
but
that's
kind
of
the
idea
that
we
have
some
neurons,
which
has
highly
non-linear
Behavior
and
the
synapse
can
adapt
to
time.
D
We
have
some
plasticity
and
learn
based
on
experience,
so
that
was
the
original
idea
behind
neural
network.
Now
we
can
learn
from
them
now.
One
interesting
difference
between
the
current
revolution
in
artificial
neural
network
and
the
way
brain
works
is
that
brain
does
not
produce
real
valued
out.
So
let's
say
you
are
using
a
sigmoid
function
or
a
real
function
in
the
artificial
Network.
Then
each
Newton's
output
can
be
anything
let's
say,
between
0
and
1,
or
between
minus
one
and
plus
one,
but
it's
a
continuous
relationship,
infinite
possibilities,
human
brains.
D
D
D
So
let's
give
an
example
with
so
this
is
from
the
speaker
as
well,
so
they
are
showing
a
nice
example
that
you
give
a
input
to
the
Deep
neural
network
of
modern
day.
So
there
are
a
lot
of
layers
hidden
layers
and
each
of
them
has
certain
functions.
So
initially,
let's
say
we
extract
the
low
level
features,
the
mid-level
features
and
the
high
level
features,
and
eventually
you
make
the
correction
normal
current
days.
State
of
the
art
state
of
the
art
networks
are
huge.
A
D
Very
accurate,
they
are
also
pretty
fast
nowadays,
but
the
training
cost
is,
if
you
think,
about
the
energy
computation.
That
is
significant,
also
the
inference
we
actually
want
to
deploy
this
thing.
You
have
to
actually
have
these
weights
and
all
of
them
are
doing
some
competitions.
There
is
a
significant.
D
The
performance
is
great.
The
main
thing
is
that
we
have
a
good
algorithm
for
learning,
so
the
original
deck
propagation
are
algorithm
from
Google,
heart
and
Hinton
and
of
course
it
has
been
updated
and
improved
for
since
then.
So
that
means
that,
because
our
neurons
is
a
abstraction
from
actual
neurons,
so
the
original
people
they
were
inspired
from
neural
network.
That's
why
it's
deep
neurometer,
but
their
neurons
is
different.
Why
is
this
not
spiking?
Why
it
is
a
sigmoid
function
or
a
value
function
because
they
wanted
to?
D
They
want
a
non-linearity,
but
not
completely
discontinence
like
zero
to
one
abrupt,
discontinuous
non-linearity,
this
kind
of
hard
non-linearity.
If
you
have
it,
it's
not
differentiable
and
the
whole
gradient
descent
depends
on
the
fact
that
the
cost
function
objective.
You
have
to
be
this
differential
so
because
we
have
a
good
algorithm.
D
D
That's
a
big
challenge
in
spiking
the
one
because
the
spiking
work,
our
mathematics
is
based
on
calculus
and
interior
algebra
that
we
are
using
here
and
especially
calculus,
which
is
very
fundamental.
So
we
we
describe
the
world
it
is.
It
works
very
well
with
this
kind
of
continuous
function.
So
when
we
face
this
kind
of
discontinuity,
our
map
I
mean
we
struggle
like
how
to
explain
that
right.
So
the
algorithm,
we
still
don't
have
a
solid
algorithm
that
we
can
use
to
strain
spike
in
neural
network,
which
is
a
huge
challenge.
D
So
this
is
a
picture
of
from
the
same
paper,
the
Silicon
Computing
ecosystem.
We
have
talked
about
this
a
little
bit
so
in
the
lower
level
we
have
the
most
transistor
build
them,
make
a
CPU
engine.
There
is
a
memory
hierarchy,
and
then
you
have
maybe
cloud-based
servers
or
CPU,
and
you
build
on
this
thing
now.
What
is
the
difference
between
what?
What
is
the
difference
between
this
thing
and
the
thing
so
first
of
all,
is
that
segregation
of
computation
and
the
memory,
or
even
bottleneck
and
in
brain,
the
synapses
of
neuron.
D
You
know
they
are
collocorate.
Computing
instruments,
so
that's
the
big
difference.
We've
already
mentioned
this.
Another
thing
is
that
in
brain
there
is
a
three-dimensional
connectivity,
because
it's
a
three-dimensional
structure
in
conventional
silicon
connections
are
in
two
Dimensions,
so
it's
a
plane
and
then
there
are
multiple
planes.
So
in
modern
I
professors
there
might
be
10
or
100
metal
layers.
But
you
know
each
of
them
are
two
layer
and
then
you
interconnect
them
through
higher.
D
It's
not
possible.
At
least
we
still
don't
have
the
technology
to
create
completely
3D
connection,
sparse,
arbitrary
connection
through
Light
plane,
I'm,
sorry,
something
very
surprising
about
brain.
So
when
you
do
plsi
connections
like
this,
what
do
you
want
is
that,
okay,
if
there
is
a
particular
computation
block-
and
there
is
another
computational
block-
we
want
to
keep
them
close
to
each
other
right
and
then
you
connect
them.
We
don't
intentionally
put
a
two
connected
block
far
apart,
but
that's
what
is
happening
in
real
brain
again.
D
D
The
far
apart
neurons
sometimes
come
back
to
the
original
neuron,
and
even
if
the
distance
increases
yes
collections
decreases,
but
it
does
not
decrease
exponential,
as
we
would
expect
it
export.
It
uses
a
power
alone,
so
here
it
is
decreases,
but
there
is
always
a
certain
bit
of
recurrent
connection
that
goes
long,
distance
away,
furniture,
which
is
not
good
for
our
way
of
computation,
because,
fundamentally
we
don't
use
time
in
our
computation
we
make
computation.
Instead,
we
know
a
input.
We
produce
another
output
time
we
explicitly
program
time.
D
D
The
computation
is
not
in
the
value
of
this
type.
The
spike
is
only
one
ordin,
only
two
bits
the
computation
is
done.
The
interval
between
this
one,
so
the
timing
interval
you
can
see
which
is
fundamentally
different
from
the
way
we
perceive,
or
at
least
we
have
been
doing
computation
so
so
that
creates
a
lot
of
architectural
and
algorithmic
challenges
there.
So
we
have
not
yet
figured
out,
obviously
how
to
involve
in
complicated
system,
but
also
is
the
challenge
for
neuromorphic
engineers.
How
can
we
design
a
algorithm
that
will
work
with
this
kind
of
system?
D
Another
problem
is
that
brain
is
inherently
stochastic
again.
What
we
think
in
our
world
is
a
problem
actually
is
a
string
in
brain,
so
in
digital
computer.
Originally,
people
use
analog
circuit
operating
systems
to
compute
things,
but
every
eventually
everybody
moves
towards
Regional.
Why
analog
is,
of
course,
more
computational
powerful?
The
problem
is:
noise.
Analog
is
not
as
immune
to
noise
as
digital
systems,
So
So,
eventually
everybody
you
know
most
of
the
technology
is
now
digital
computation
is
at
least
so.
We
think
of
stochasticity
anything
random
as
bad.
D
D
It
stands
out
and
we
are
not
yet
sure
how
it
is
actually
happening,
but
it
turns
out
that
stochasticity
is
actually
a
feature
in
human
brain.
At
least
it
turns
out
that
that's
how
that's
why
Evolution
has
done
it.
This
way
that
yes,
brain
computations
are
not
exact.
If
you
give
the
same
input
twice,
you'll,
probably
not
get
the
same
output
wise
as
we
want
in
our
digital
system,
but
apparently
it's
a
good
thing.
D
So
what
some
of
the
cognitive
neuroscientists
and
computational
model
is
to
think
is
that
probably
this
helps
them
to
not
be
stuck
to
a
local
maximum.
So
that's
a
big
problem
in
training
algorithms.
You
know
familiar
with
the
more
machine
learning
one.
If
you
are
stuck
in
a
certain
place,
then
suddenly
there
will
be
a
random
event
that
will
happen
and
it
will
move
you
ever
you
are
not
supposed
to,
but
it
will
just
randomly
move
it
away
and
it
actually
helps
you
to
move
in
otherwise,
you'll
be
stuck
and
circling
inside.
D
So
somehow
again,
all
the
details
are
not
clear
yet,
but
there
have
been
some
very
interesting
one
that
the
stochasticity
is
leveraged
in
brain
as
if
it's
a
feature
so
so
far,
we
have
been
using
silicon
Computing
platform
to
do
the
Deep
learning
thing.
So
the
huge
energy
and
speed
requirement
produces
a
challenge,
so
we
are
trying
to
now
do
neuromorphic
Computing
with
all
the
advantages
of
the
human
brain.
D
So
let
me
just
show
a
little
bit
of
timeline
again
all
the
details-
I'm
not
going
to
go
towards
this,
but
in
1940s
to
1960s
people.
First
idea
of
what
is
a
spiking
neuron
Hodgson
actually
had
a
famous
model
of
neuron,
which
is
the
very
realistic
model
and
also
during
the
same
type.
In
the
Silicon
world.
We
have
been
making
big
breakthrough.
The
transistors.
D
1970S
1980s
was
also
great
time
for
artificial
intelligence.
He
started
greatly
so
there
was
the
original
ideas,
but
then
the
something
called
the
AI
winter
clean
and
all
that,
but
I'm
not
going
there.
But
the
main
idea
is
that
yeah
in
the
1980s
they
had
the
direct
propagation
learning
algorithm,
which
will
become
very
powerful
later
and
also
people
came
up
with
the
idea
of
neuromorphic
competing.
This
was
actually
led
by
karval
meat
at
Caltech.
D
He
wanted
to
build
analog,
Computing
blocks
for
neurons
and
sea
labs.
In
fact,
he
was
one
of
the
pioneers
of
vlsa
as
well,
so
he
first
started
with
blsi
the
pioneering
work
wrote
the
first
textbook,
I
think
and
then
in
1980s.
He
also
realized
that
it
has
certain
limitations,
it's
great
for
what
it
does.
It
has
certain
limitations,
so
he
started
exploring
these
brain
like
computation
Paradigm
of
computing.
He
named
this
community
also
a
little
unnoticed
work
in
1971,
Leon
Chua,
a
theoretician
in
University
of
California.
D
He
proposed
that
there
is
actually
a
fourth
fundamental
element
from
exhibited
principle
or
who
should
exist
called
Memories,
and
he
developed
this
in
1976
with
his
student
as
well.
So
this
idea
was
a
theoretical
work.
Nobody
seen
this
year.
People
did
not
care
that
heard
that
much
about
it
to
be
honest,
so
it
was
there
in
1970s.
It
just
happened
in
1971
and
then
in
1990s.
You
know
gpus
formally
introduced
address
event
recording
this
thing
came
for
once
it
is
by
communication.
Some
of
the
spiking
models
were
introduced.
D
So
many
there
are
many
others
present
challenges
in
neuromorphic
world,
so
I
am
just
going
to
and
put
this
in
the
slide.
So
if
you
want
to
explore
any
of
them,
you
can
do
that.
But
let
me
go
into
the
current
research
and
especially
like
the
three
generations
of
neural
network,
so
to
make
it
transition
that
what
happened
so
in
the
1940s
and
50s,
the
first
idea
of
neural
network
was
done
with
perceptron.
D
A
D
D
So
these
deep
plates
are
all
continuation
of
the
second
generation
of
neurons
and
the
third
generations
of
memory.
Is
these
spiking
neurons,
which
are
more
grainy?
So
there
are
quite
a
few
varieties.
I
will
show
one
at
least
so,
for
example
in
spiking
neural
network,
this
third
generation
company-
by
the
way
this
is
not
my
nomenclature.
This
is
from
a
seminal
prepared
by
mass
I'll,
put
the
link
here
as
well,
so
he
was
saying
that
this
third
generation
spiking
neural
network.
So
let's
look
at
this
example.
D
So
inputs
are
not
continuous
tension.
It's
either
one
or
zero
information
is
actually
encoded
in
the
time
interval
between
inputs.
So
let's
say
this
is
the
input
layer
and
the
inputs
are
coming
and
thus
neurons
are
producing
some
output.
The
input
results.
They
are
going
through
a
bunch
of
Weights,
so
input
one
is
going
through
weight.
One
W
one
input
2
is
going
to
W2.
You
can
think
of
this
W1
W2
w3s,
synaptic
string.
You
can
learn
these.
You
can
modify
these
and
based
on
that.
D
It
determines
how
much
it
will
contribute
to
the
next
name.
So
in
the
literature
they
will
call
the
first
layer,
pre
neuro
and
the
second
layer
postnode,
and
the
connection
is
the
synapse,
so
every
synapse
will
have
a
pre-neuron
and
a
posterior.
Now
in
the
neuron,
the
operation
that
is
happening
is
first
of
all,
a
summing
operation.
All
the
inputs
are
sum
and
then
a
thresholding
operation.
You
make
a
decision,
so
there
is
a
threshold
value.
D
If
you
go
behind
this
ratio,
this
post
neuron
produces
an
output
Spike
and
that
may
connect
to
the
next
layer
of
view
right,
and
this
is
the
feed
forward
architecture,
so
inputs
to
outputs
left
to
right,
but
in
actual
neural
network
in
human
brain.
There
is
a
lot
of
recurrent
connection
again,
most
of
the
Revolution
and
the
learning
method
that
has
developed
artificial
irrele
method.
D
But
let
me
show
a
little
bit
of
how
a
neuron
looks
like
so
when
a
spike
comes
up,
the
membrane
potential
of
the
neuron
goes
up,
and
then
there
is
no
spike.
It
goes
down
the
case
and
then
again
a
spike
camps,
and
it
goes
up,
goes
down
and
so
forth.
Once
there
is
enough
number
of
spikes
come
in
a
quick
interval,
you
can
reach
the
threshold
and
once
you
reach
the
thresher
voila,
you
have
the
green
post
space.
So
that's
kind
of
the
idea.
D
This
is
what
is
happening
after
Spike
happens,
neuron
kind
of
goes
through
a
sleep
mode
for
a
while
it's
called
the
refractory
period.
So
during
this
time,
even
if
input
signal
comes,
it
does
not
integrate
the
inducing.
It
does
not
care.
It's
just
incomplete
hibernation.
After
this
reflectory
period
it
again
start
integrating.
D
You
can
think
of
in
human
as
well,
so
if
you
take
give
so
right
now,
I'm
talking
for
a
long
time,
so,
probably
initially,
you
are
kind
of
going
a
little
bit
listening
to
getting
into
the
step.
But
after
a
while
you
are
now
to
have
been,
you
know
getting
like
how
long
will
this
continue?
So
this
is
your
refractory
period.
So
at
this
time
you
are
not
taking
in
any
input.
You
are
probably
zooming
out
and
then
maybe
after
five
minutes
since
we're
like.
D
Oh,
maybe
I
can
listen
again
and
then
you
get
interested
again.
So
what
spiking
neural
network
people
are
trying
to
do
is
that
they
are
trying
to
see
that.
Oh,
we
know
from
cognitive
neuroscientists
that
human
brains
neurons
have
this
functionality
in
artificial,
neural
network.
We
are
not
doing
any
of
this.
We
have
a
simple
non-linearity,
continuous
function.
Do
differentiation,
gradient
descent,
learn
this
thing
work
great,
but
maybe
the
recent
Evolution
came
up
with
this
is
maybe
all
these
things
are
not
just
there.
For
you
know
random
purpose.
D
B
D
D
The
next
10
minutes
I'm
going
to
wrap
it
up.
So
basically,
what
are
the
challenges
and
opportunities
in
this
field?
I
am
going
to
mention
them,
so
snls
are
behind
the
Deep
Learning
Network.
To
be
honest,
like
most
learning
tests,
spiky
neural
network
is
still
not
as
accurate
as
current
state
of
priority.
Okay,
so
that's
one
problem.
Another
problem
is
the
data
available
most
of
the
data
at
the
form
they
are.
D
D
So
one
way
people
have
been
tying
is
that
people
are
doing
conversion
so
first
they
do
a
deep,
Learning,
Network
training.
Then
they
convert
the
Deep
neural
network
to
spiking
New
London,
where
the
inputs
are
converted,
increase,
spikes
and
outputs
are
very
interesting.
There
are
rescaling
and
normalizing
method
that
people
have
been
developing
so
far.
This
has
given
the
best
result.
The
problem
is
that,
of
course,
this
is
not
the
real
deal,
because
this
is
not.
D
Element
we
are
kind
of
avoiding
the
problem.
Also.
The
reason
we
want
to
do
spec
in
your
network
is
to
avoid
that
energy
cost
and
high
speed
and
good
latency.
Unfortunately,
this
whole
conversion
mechanism
solves
the
problem,
but
then
the
final
network
does
not
have
those
Energy
Efficiency
anymore.
I
mean
they're
much
moves
compared
to
what
is
spiking
can
do
so
it's
a
way
around
the
problem.
I,
don't
think
it's
actually
solving
the
problem,
but
it's
it's
a
it's
a
very
interesting
thing.
People
have
been
doing
having
some
good
results.
D
What
would
like
to
do?
The
community
I
think
would
that
actually
Spike
based
learning
algorithm
and
then
there
can
be
supervised
or
there
can
be
unsupervised
learning,
as
you
know
so,
supervised
learning.
The
problem
is
the
gradient
reason.
So
what
people
have
done
is
that
they
have
approximated
hard
discontinuity
with
something
approximating
continuity,
so
that
it's
good
enough
and
then
you
can
do
calculus
and
great
ideas.
D
There
have
been
quite
a
few
other
works,
but
I'm
not
going
to
happen
now
for
unsupervised
learning.
It
has
huge
potential
because
in
starting
from
1940
is
the
50s
Neuroscience
is
new,
so
there
is
a
rule.
Called
heaviest
flow
is
neurons
that
fire
together,
so
in
our
previous
figure.
If
you
look
at
this
thing
that
if
the
post
neuron
spikes
after
the
pre-neuron,
then
we
say
that
the
weight
should
increase.
So
we
we
want
some
influence
from
input
to
pre
neuron
to
post
neuron.
D
It
looks
like
they
should
connect
together
so
neurons
that
fire
together
should
work
together.
So
if
the
delta
T
the
time
difference
between
post
entry,
neuron
is
positive,
we
increase
the
weight.
So
that's
the
learning.
How
do
you
change
the
update
and
if
it
is
the
reverse,
then
you
decrease
the
weight.
So
it's
shown
here
in
the
field.
So
if
the
delta
T
is
positive,
increase
the
weight,
if
your
Delta
is
negative,
you
decrease
the
weight
and
it
exponentially
boost
and
people
have
used
these
unsupervised.
D
Learning
do
listen,
local
learning,
so
just
you
don't
need
the
final
full
theme
and
shown
some
good
results
for
every
data.
It's
a
very
promising
field.
Still
there
is
problem
because
local
learning
rule
how
to
extend
it
to
a
the
full
network
is
not
clear
here:
electricity,
so
learning
algorithm
development
is
a
huge
challenge
in
this
field,
for
there
are
a
few
other
things
which
are
both
challenges
and
also
because
of
challenge.
D
They
are
actually
opportunity,
so
we
can
go
beyond
static
image
classification,
so
right
now,
most
of
the
work
in
the
Deep
neural
network
quality
is
static,
image
classification
or
that
kind
of
thing
nowadays,
lateral
language
processing
as
well
so
most
SSM
work
has
been
benchmarked
with
static
image
classification.
Can
we
do
more
than
that?
So
there
are
now
even
based
camera,
where
the
image
is
not
trained,
based,
it's
small
event
based
which
is
naturally
suited
to
how
human
brain
processes
think
so
people
are
doing
actually
very
interesting
research
with
even
based
camera
and
neuromorphic.
D
They
seem
like
a
perfect
match,
so
I
am
hoping
that
there
will
be
interesting,
a
lifelong
learning
to
human
beings
life
over
a
long
time.
So
that's
a
big
challenge,
maybe
neural
network
spec
neural
network
can
help
with
that.
We
also
forget
things,
and
there
has
been
some
interesting
work
in
neural
network.
D
It's
SNL
feel
that
how
forgetting
is
actually
helpful
and
how
can
you
avoid
what
is
called
catastrophic,
forgetting
how
can
you
elaborate
with
your
data,
so
this
has
been
a
big
problem
when
the
Pioneers
like
interns-
and
they
all
talk
about
that
when
a
child
learns
they
learn.
So
our
chair,
GPT,
is
learning
language
and
doing
some
apparently
good
results,
but
it
is
trading
on
a
huge
amount
of
data
set
huge
competition,
but
unlikely
that
a
human
child
learns
a
language
with
so
much
label
data.
It
definitely
doesn't
so.
D
Obviously
something
else
is
going
on
learning
with
fewer
data.
Can
we
can
we
do
something
better,
then
Neuroscience
connection?
This
is
purely
science
oriented
people
that
can
we
learn
more
about
your
human
brain.
Of
course
that's
fascinating
in
its
own
right,
even
if
does
not
have
any
computational
Hardware
development
with
Device
circuit
who
design.
So
this
unlisted
the
building
block
was
the
revolution.
Can
we
come
up
with
new
circuits
and
device
that
are
perfectly
suitable
for
this
neural
network
or
implementation?
So.
D
Magnetic
Ram
people
have
been
promoting
this
emerging
devices.
They
are
perfectly
suited
for
this
new
paradigm,
big
brain
Chiefs
like
internal
IBM,
true
north,
and
all
this
thing
people
have
developing
so
billions
of
new
or
millions
of
neuron
videos
have
seen
absence
asynchronous
address
event,
representation
or
AER.
So
our
conventional
computers
are
digital.
They
have
a
synchronous.
D
Synchronous
means
that
it
has
a
global
clock
and
every
clock
cycle
it
doesn't,
but
human
brain
does
not
work
like
there
is
not
a
global
clock
in
our
brain,
so
it
is
local
calculation
and
there
is
asynchronous
communication
whenever
event
happened,
it
is.
There
is
something
so
in
neurology
Computing.
What
we
ideally
want
to
do
is
not
have
a
global
curve.
Instead
event
based
communication,
there
will
be
sparse
connection
and
much
less
energy.
D
D
Design
is
much
more
difficult.
There
is
the
challenge
and
as
well
as
the
opportunity,
a
network
on
cheap.
How
can
you
connect
them?
People
have
been
proposing
interesting
way
to
do
the
communication.
In
fact,
there
has
been
some
exciting
who
are
using
optical
electronic
calculator,
so
the
computational
done
on
Silicon
or
maybe
superconducting
devices
or
the
nist
is
doing
this.
Actually
so
some
highly
computationally
efficient
platform.
But
then
the
communication
data
communication
will
not
be
through
wires,
but
instead
through
Auto
Electric
example,
photonics
thing,
so
amazing
things
can
happen
in
that
field.
D
I
I
personally
think.
The
main
goal
is
to
go
beyond
volume
and
Computing,
so
we'll
do
processing
and
compute
in
the
same
place.
So
no
bottleneck
of
that
memory,
bottleneck,
non-volatile
technology
have
been
revolutionary,
so
I've
talked
about
name
Lister,
so
2008,
first
experimental
demonstration
of
membership,
and
since
then,
in
the
last
10
years,
10
15
years,
there
have
been
plenty
of
groups
who
have
proposed
new
devices,
still
not
matured,
but
hopefully
we'll
go
there
pretty
soon.
And
then
there
is
the
mixed
signal:
analog
Computing.
D
How
can
you
build
good,
neurons
and
synapses?
That
can
do
this
thing:
a
memory
stick
dot
product
engine,
so
the
fundamental
linear,
algebra
operation
in
this
thing
is
Matrix
Factor
multiplication.
So
you
can
build
cross
person
and
visitors
and
do
this
using
physics.
So
what
actually
happens
is
that
Ohm's
law
does
the
multiplication
and
kashif's
current
load
does
the
addition,
so
you
don't
have
to
explicitly
do
multiplication
radiation,
like
you
have
to
do
in
a
digital
computer,
the
inherent
device
remix
of
this
device
through
this
computation.
D
For
you,
the
problem
is
their
analog,
so
not
as
accurate
or
robust
as
the
digital
region.
So
then
you
need
to
do
algorithm,
Hardware,
poor
design,
so
that,
even
when
there
is
variation
in
your
analog
circuit,
your
algorithm
will
require,
which
is
a
huge
challenge.
People
are
working
on.
Another
thing
is
stochasticity
probabilistic,
so
there
has
been
Boseman
machine
offering
Network
people
have
been
trying
to
see
the
probabilistic.
How
can
he
include
probability
and
stochasticity
into
algorithms?
D
D
B
E
Yeah
I
have
a
quick
question
about
the
snn
yeah,
so
I
see
that
you
have
the
the
weights
that
are
actually
being
multiplied
by
the
input
swipe
train
and
you
sum
the
results
of
those
weights.
So
there
must
be
a
backward
propagation
process
for
supervised
learning
process.
Pretty
straightforward!
You,
you
figure
out
how
well
your
prediction
did
based
on
your
threshold
to
the
expected
output.
E
D
So
there
is
this
post
neuron
and
the
pre
neuron
if
the
post
neuron
fires
after
the
clean
neuron,
what
the
assumption
is
that
that
this
post
neuron
has
fired
because
of
the
contribution
from
the
P
neurons.
So
probably
the
information
that
is
coming
from
this
P
neuron
to
this
post
neuron
is
useful.
So,
let's
make
this
stronger,
make
this
weight
stronger.
D
Conversely,
I
know
I,
know
that
you
might
have
questions
about
it.
I
also
have
questions
but
I'm,
just
saying
that
what
the
algorithm
is
and
when
the
post
neuron
fires
before
the
premiere,
it's
the
reverse,
so
you
reduce
the
weight.
So
that's
how
you
do
the
weight
update.
You
are
not
looking
at
the
final
result:
There
Is,
No,
Label
data,
it's
a
local
thing.
So
in
one
layer
you
are
looking
at
the
output
and
the
input
and
based
on
that,
you
are
making
apparently
from
Neuroscience.
D
What
we
have
seen
is
that
this
famous
rules
from
heps
called
the
HEBs
will
ABN
rule
is
that
neurons
that
were
fired
together
should
wire
to
go.
So
if
this
post
neuron
fires
after
this
pineal
that
we
will
assume
that
their
connection
is
particular
width.
W3
in
this
case
should
be
potential
in
their
language.
So
that
means
the
weight
should
go
up
or
if,
in
the
converse
case
it
should
be
depressed
or
it
should
go
down.
Apparently
that's
what
happened
now.
Of
course,
that's
on
its
own
is
not
enough,
so
there
is
also
inhibitory
correction.
D
So
this
so
there
are
exciting.
I
was
not
able
to
go
over
all
the
details,
so
there
are
exciting
excitatory
connections
and
inhibitory
connections
so
that
things
done
don't
go
exponentially
up
or
exponentially
down
and
you
can
kind
of
keep
it
stable.
But
this
is
one
unsupervised
algorithm
and
there
has
been
a
work
if
I
remember
DL
and
others
2017
related,
but
they
have
shown
that
just
by
doing
HTTP
within
a
few
tweaks
unsupervised,
there
were
able
to
classify
me
status.
D
We
just
love
the
huge
data
set
of
currently
maybe,
but
they
were
able
to
get
almost
the
state
of
that
equation,
using
working
or
developing
on
this
HTTP
learning
map.
So
HTTP
is
one
very
famous
learning
method.
Some
other
unsupervised
learning
method
actually
I
have
been
working
with
is
called
short-term
plasticity.
So
in
conventional
neural
network
you
have
a
weight
right.
You,
land,
This,
Way
throughout
the
training,
maybe
create.
D
In
this
end,
once
you
have
learned
the
width,
you
put
the
width,
so
let's
say
I
have
learned
what
is
the
base
field
of
W1
for
this
problem?
You
just
program
these
devices
and
then
input
comes
you
get
the
correct
output
in
real
synapse.
That's
not
how
it
happens.
Yes,
there
is
a
long
term
memory,
it
lands
and
it
makes
the
connection,
but
there
is
also
a
short-term
memory,
so
synaptic
weight
is
not
fixed.
D
Synapse
itself
learns
through
time.
So
if
the
input
sequence
comes
fast
and
Fast,
then
synapse,
where
it
keeps
increasing
increasing,
but
then,
if
it
input
stream
becomes
slower,
the
CDF
system
goes
down.
I'll
give
a
simple
example.
So,
for
example,
when
I
am
listening
to
someone
initially
I
am
interested.
My
attention
is
here
the
this.
If
you
think
of
my
in
my
attention
as
this
synaptic
weight,
some
Connection
in
your
brain,
it
is
getting
stronger
and
stronger,
but
so
it's,
but
after
a
while
it
start
becomes
saturated.
D
So
maybe
I
have
a
Time
level
of
time
scale
of
10
15
minutes
after
10
15
minutes,
I
cannot
follow
things
getting
saturated
and
then
actually
it
becomes
depressed,
so
I.
In
fact
now
I
am
just
passing
these
two
from
one
year
to
another
right
very
crude
example,
but
I'm
just
saying
that
in
actual
human
synapse
is
not
fixed,
as
we
are
doing
in
our
artificial
neural
network,
which
is,
of
course,
they
know
this.
This
is
an
abstraction
useful
for
computation,
but
I'm
just
saying
that
actual
synapse
are
not
fixed
in
time.
D
It
changes
on
input,
that's
actually
a
part
of
my
research,
so
the
particular
device
that
I
work
with
has
a
short-term
plasticity
which
only
works
on
the
input.
It
doesn't
even
look
at
the
immediate
output
forget
the
final
round
like
no
supervision
at
all,
like
this
completely
unsupervised.
Thing
apparently
is
helpful
for
online
learning,
so
it
helps
human
being
to
process
things
all
animals,
so
our
synapses
in
organic
organisms,
the
synapses,
the
neurons
in
organisms,
have
quite
a
few
of
these
unsupervised
learning
mechanism
that
apparently
help
them
to
add
up
to
new
sarcomasters.
D
So
the
actual
algorithm
of
why
this
is
computational
helpful
is
not
clear
here
so,
and
there
is
a
particular
egg
problem,
so
computationally
simulating
human
brain
on
conventional
volume
and
computer
is
very
difficult.
You
have
to
build
that
good,
neuromorphic
computer
spiking
neural
computer
to
simulate
human
brain
to
understand
things,
but
to
have
to
build
a
good,
striking
neural
network
based
Computing
system.
You
need
to
understand
how
human
brain
does
this.
E
C
D
These
two
yeah,
the
main
difference
I
mean
there
are
quite
a
few
differences,
but
the
main
difference
is
that
the
data
format,
so
in
this
case
data
can
be
either
a
spike
or
notice
right,
but
in
terms
of
floating
services
or
whatever
CPU
you're
using
in
that
case,
data
is
represented.
As
a
continuous
number,
you
can
take
a
range
of
number
between
0
to
1
right,
it's
a
continuous
value,
that's
what
part
is
Spike
is
Spike
is
either
Spike
or
not.
D
Yeah,
so
that's
one
method
that
I
mentioned
that
one
of
the
way
people
are
doing
this
is
they
are
doing
conversion
business.
So
first,
let's
solve
the
problem
with
deep
user
Network
what
you
are
saying
and
then
they
convert
these
to
a
equivalent
spiking
neural
network.
Now
it's
not
as
simple
as
its
sounds,
because
there
are
also
fundamental
differences
between
the
neurons
and
synapses
between
artificial
neural
network
and
stacking
neural
network
synapses.
As
I'm
saying,
synapses
is
much
more
functionality,
real
synapses,
but,
most
importantly
neuron.
D
Neuron,
these
are
three
a
three-dimensional
ordinary
differential
equation
with
three
state
variables
are
even
more
than
that,
so
pretty
complex
Dynamics
with
a
very
discontinuous
thresholding,
which
is
the
problem
for
the
calculus
and
radialism,
so
converting
from
this
neural
lateral
energy
except
through
superficially.
It
looks
very
similar
because
of
this
two
main
difference
that
the
neuron
block
is
very
different.
D
However,
it
does
not
take
the
best
advantage
of
spiking
more
than
that,
so
the
final
Network
that
you
get
it
actually
produces
good
accuracy,
but
the
main
promise
of
spagging
neural
network
is
that
low
energy,
patient
computation
these
circuits
are
actually
becomes
huge
and
consume
a
lot
of
power.
So
I
think
personally,
people
might
be
thinking.
I
personally
think
that
only
now
our
goal
should
be
to
come
up
with
an
algorithm.
That
is
is
a
difficult
problem
that
required
algorithm,
which
is
for
Spike
based
and
it
can
be
supervised.
It
can
be
unsupervised.
D
C
D
No,
no
so,
personally
speaking,
I,
don't
think
that
it
will
replace
all
computation
I
still
think
that
what
we
have
done
in
digital
computer.
It
is
actually
very
good
for
certain
things.
So
if
you
are
doing
math
operation
right,
I
think
you
actually
want
deterministic
always
same
input,
same
output,
that
kind
of
one
right-
if
you
don't
want
to
discussing
that,
but
there
are
lots
of
real
life
activities.
So
if
you
want
to
build
general
intelligence
right,
so
something
which
is
artificial,
general
intelligence,
that
is
which
is
kind
of
the
Holy
Grail
of
article.
D
D
Nobody
predicted
going
to
mood
and
stuff
like
that.
Moving
to
Moon
seemed
like
a
very
difficult
problem
and
moving
like
human
beings
talking
like
human
beings
like
the
easiest
program,
because
we
do
this
without
any
effort.
I
think
this
is
very
misleading.
Actually
speaking,
language
perception.
These
are
very
difficult
things.
We
do
it
effortlessly,
because
Evolution
has
prepared
us
millions
of
years
to
do
this
effortlessly
when.
B
D
You
actually
went
there
and
try
to
build
the
machine
that
can
do
the
same
things
in
intelligent
robot.
We
found
out
that
it's
such
a
difficult
problem
to
build
this
thing,
so
sometimes
a
problem
that
can
seem
easy
is
actually
very
difficult
and
sometimes
so
that
kind
of
thing
happens.
So
nowadays,
in
the
palm
of
my
hand,
I
can
have
a
smartphone
which
is
more
computational
power
than
the
computer
that
was
used
to
send
a
rocket
to
the
moon.
D
You
know
and
is
still
now
we
don't
have
a
robot
that
can
talk
like
human
beings
and
move
like
human
beings
are
as
flexible
and
as
they
are
going
to
be,
which
seemed
like
people
would
like.
So
if
you
read
like
the
movies
famous
movie
1960
or
the
2001
Space
odyss,
so
let's
retail,
around
1940s
and
50s,
it
was
long
future
2001
in
Space
Odyssey,
so
by
2000.
Obviously,
we'll
have
this.
We
are
now
in
2023.