►
From YouTube: Hebbian Learning in Neural Networks (Brains@Bay Meetup)
Description
https://www.meetup.com/BraIns-Bay/events/265724515/
Meeting starts at 30:00.
Welcome to another session of Brains@Bay! This time we will be talking about Hebbian Learning, the main mechanism used by the brain to learn new experiences.
As always, we will start with a perspective from neuroscience, then move on to understand how these concepts can be applied to machine learning. Our meetup is a journal club, so expect a lot of very interesting discussions during the presentations.
We've changed the location for this next event to make it more comfortable for everyone - we've partnered with UCSC and will hold the event at UCSC Silicon Valley Campus. Please see the details on the meetup page.
Agenda:
** 06:30 - 07:10: Networking and Pizza Time!
** 07:10 - 07:50: Hebbian Learning in Neuroscience - Floran Fiebig, PhD
Title: Bayesian-Hebbian learning in spike-based cortical network simulations.
Many Hebbian learning rules exist, serving a variety of local learning rules that both explain neuroscientific findings and shape various networks in machine learning. In a biophysical model of spiking cortical neural networks, we leverage abstracted processes of synaptic and nonsynaptic changes, to derive a local, online Hebbian learning rule for spiking neurons inspired by Bayesian statistics. The resulting spiking implementation of a Bayesian Confidence Propagation Neural Network (BCPNN) is capable of explaining a broad array of neuroscientific findings and perform a variety of cognitive tasks, but may also be interesting to ML investigators puzzled by the spiking nature of neuronal computation, and how one might want to approach the interesting problem of mapping existing learning rules onto a network that leverages discrete spikes as a means of information propagation and unsupervised online learning. SLIDES: https://prezi.com/view/Hf1MBr8cfXSNs8LeVDqt/
** 07:50 - 08:30: Hebbian Learning in Machine Learning - Thomas Miconi, PhD
Title: Differentiable plasticity: training self-modifying networks for fast learning with Hebbian plasticity
Neural networks are usually trained over long periods with massive amounts of data. By contrast, animals and humans can learn complex information very quickly. This ability for fast lifelong learning results from (largely Hebbian) synaptic plasticity, carefully tuned by evolution. Here we show that plastic networks can be trained to learn quickly and efficiently, using standard gradient descent instead of evolution. The trained networks can quickly learn high-dimensional information and use it to solve a task. Our results show that differentiable plasticity may provide a powerful novel approach to the learning-to-learn problem.
** 08:30 - 09:00 Discussions and Wrap-up
0:00 Starting soon...
30:51 Floran Fiebig - Hebbian Learning in Neuroscience
1:20:00 Thomas Miconi - Hebbian Learning in Machine Learning
A
B
B
C
B
B
B
D
C
C
C
C
C
C
C
G
F
G
C
A
G
G
Have
two
thoughts
for
today:
that's
Thomas
over
there
and
Lauren.
So
thanks
a
lot
for
coming,
so
we're
experimenting
different
formats
and
we
decided
to
just
shoot
oxygen
to
your
4.
So
we
have
more
time
for
discussion
and
that's
it.
So
we
start
with
Lauren
I
miss
elbow
and
introduction
for
you,
no
matter.
E
G
F
F
So
let's
go
the
edge
a
little
bit
on
both
of
these
topic,
which
means
I'm
going
to
ride
roughshod
over
both
the
neuroscience
and
the
machine
learning
point
just
in
order
to
build
a
bigger
arc
that
actually
rose
in
everybody
into
this
conversation.
So
let's
see
how
that
goes.
So,
there's
a
couple
of
things.
I
want
to
address
today,
so
obviously
we're
going
to
talk
about
heavy
running
and
what
it
can
do,
in
particular
what
we
can
do
for
a
Bayesian
inference
and
all
might
one
implements
of
in
my
place.
F
F
So
how
do
we
simulate
these
things
in
a
defensible
way
and
what
might
be
learned
from
all
of
that?
So
just
as
a
quick
reminder
right.
So
this
is
overly
broad.
Obviously,
but
if
we're
looking
at
machine
learning
additional
networks
in
neurobiology,
we
see
that
you
know
the
network's
going
like
these
connected
know
forward
ways,
whereas
neurology
is
highly
current.
F
You
know
an
answer
typically
clocked
or
trained
and
crashes
and
where's.
The
brain
is
completely
asynchronous.
No
neuron
waits
for
another
neuron
to
do
its
thing.
Writers
I
run
all
the
time,
and
so
the
activity
is
in
spike
and
bursts
and
oscillations
of
whatnot,
as
opposed
to
a
big
vectors
that
are
being
passed
around
and
you're
running
continuously,
instead
of
like
static
data
sets,
and
you
have
to
run
on
the
basis
of
these
spike
traces
as
opposed
to
real
valued
gradients.
F
F
F
Being
passed
around
or
bits
they
communicate
with
spikes.
Spikes
are
like
these
events
v,
so
just
about
a
millisecond
activation
peak,
half-width,
so
they're
very
fast
events
and
that
sort
of
discrete
things,
and
then,
on
top
of
that,
when
you
look
at
real
activity,
it's
not
just
individual
spikes.
They
often
come
in
like
these.
These
first
packets
here,
where,
like
many
spikes,
sort
of
bunched
together
and
these
gamma
bursts,
and
then
you
might
get
multiple
of
those.
F
F
F
These
vesicles
are
filled
with
neurotransmitter
and
with
the
help
of
two
different
proteins,
that
kind
of
bind
to
the
membrane
and
then
the
calcium
can
buckle
those
proteins
in
order
to
release
the
content
of
their
neurotransmitter
inside
vesicles
into
the
synaptic
cleft,
and
about
half
a
millisecond
later
like
these
there's,
no
transmitter
will
bind
at
the
postsynaptic
side
and
open
these
channels
for
sodium
ions
to
flow
in.
So
now
you
have
a
map
massive
influx
of
sodium
ions,
which
is
going
to
change
the
postsynaptic
potential.
F
So
the
membrane
potential
on
the
other
side
right
on
the
receiving
neuron,
is
going
to
change
as
a
result
of
this
pulse
train
up
here.
So
if
we
take
a
look
at
what
that
actually
looks
like
so
here,
you
have
us
I
train
brightness,
like
a
brief,
come
across
here,
so
an
80
Hertz
packet
on
that
four
spikes
and.
E
F
E
F
Packets
here
is,
you
know,
maybe
1
to
2
point
5
millivolts,
maybe
right
so
you
cannot
leave
many
synapses
to
cooperate
and
you
don't
need
to
spike
trains
as
opposed
to
single
spikes,
to
really
take
off
a
person
and
acknowledge
it
higher
all
right
that
being
a
quick
recapitulation
of
biology.
Let's
move
on
to
some
other
machine
learning
part
of
this
heavier
money.
Most
of
you
will
be
familiar
with
Bayes
rule,
just
a
simple
idea
that
the
probability
of
conditional
event
can
be
inverted.
F
F
F
F
Different
attributes
X
that
you're
using
here
are
actually
independent
of
one
another,
or
at
least
independent
enough
such
that
you
can
apply
this
rule
and
then
the
third
problem
is
that
often
times
when
we
ability
specifies
these
are
continuous
variables.
And
so
how
are
you
going
to
do
this
now,
with
sort
of
some
probability
on
some
action
being
true
or
false?
So
it
cannot
need
to
have
a
way
to
map
those.
We
will
return
to
all
those
problems
before
we
do.
F
F
E
F
F
F
We're
saying:
okay,
I'm
going
to
have
a
pool
of
inhibitory,
neurons
pools
of
connective
neurons
and
I'll,
organize
them
in
such
a
way
that
different
intervals
of
the
spectrum
of
the
attribute
I
want
to
code
here
are
reflected
on
these
different
mini
columns,
so
these
subunits
and
then
they
can
all
inhibit
each
other
and
compete
for
activation
of
inhibitory.
Neurons
such
that
when
one
of
those
activates
it's
going
to
shut
down
all
the
other
ones,
which
gives
you
a
very
nice
information
contrast
similar
to
what
we
observe
individual
cortex
and
so
keep
in
mind.
F
E
F
E
F
F
We
can
map
now
these
continuous
problem.
We
otto's
also
get
automatically
added
correlation
of
the
different
inputs
through
the
inhibition,
meaning
there's
going
to
be
winning
mini
column
if
1
mini
column
strongly
enough
to
say,
I
know
what
this
is,
that
it's
gonna
shut
down
all
the
other
ones,
which
gives
you
nicely
independent
input
parameters
if
you
wanted
to
now
processing
with
these
kinds
of
Units.
So
now
that
again,
two
machine
line-
and
so
you
want
to
use
these
these
kinds
of
organized
structures
down
so
nest.
Hyper
comes
in.
E
F
F
And
you
already
know
how
to
build
your
biases
and
your
weights
in
this
logarithmic
expression
in
order
to
make
that
system
work.
The
one
thing
that
you
now
need
to
do,
of
course,
which
is
a
bit
of
a
complication,
is
you
need
to
have
an
exponential
for
the
output
because
you're
working
in
logarithmic
space
and
you
need
to
renormalize
the
output
from.
B
F
All
the
different
inputs,
which
nicely
also
solves
the
problem,
one,
namely
that
you
can
now
have
unobservable
variables
and
that's
not
really
a
problem
because
of
course,
in
the
denominator
here,
all
the
all
the
XS
right
there
get
some
producted
up
without
the
logarithm,
but
are
the
same
for
all
classes.
So
the
order
of
the
different
probability
stays
the
same
and
once
you've
been
organized.
F
You're
gonna
have
one
winning
pattern
so,
but
we
now
need
to
have
a
running
estimate
of
the
prior
probability:
X
the
posterior,
the
probability
wine
and
the
coagulation
the
probability
x
and
y
active
together.
So
how
can
you
do
live
in
a
biologically
defensive
way
in
a
way
that
we
can
defend
right
as
possible?
And
so
again
now
I
turn
back
to
2
pi
ology.
So
we
somehow
need
to
have
a
local
site
based
rule,
because
you
know
real
neural
networks
work
with
spikes,
so
we
can't
just
use.
You
know
simple,
straightforward
correlation
rules.
F
D
F
There
of
mine,
and
actually
there's
this
section
out
of
this
one,
so
one
of
those
is
that
the
units
inside
the
mini
column
are
highly
recon
themselves.
So
all
the
nodes
that
locally
selected
for
the
same
features
also
have
a
higher
than
chance
probability
of
connecting
with
one
another,
but
there's
gonna
be
recurrence
beyond
that
and
that
I
will
get
to
in
the
next
slide.
F
This
work
that
I'm
now
talking
about
is
a
work
of
a
colleague
of
mine,
Phil
Tully,
who
did
his
PhD
just
a
year
or
two
before
me
in
Stockholm
too
far
away
to
nicely
map
such
a
Bayesian
running
rule
on
to
spiking
neural
networks.
So
you
now
have
two
spike
trains
of
connected
norm:
pair
I&J,
so
pre
and
the
post
synaptic
neuron.
Any.
F
You're
gonna
do
now
something
that
is
super
common,
which
you've
already
seen
that
these
neurons
do,
namely
they
have
some
kind
of
a
temporal
filter
to
to
aggregate
the
spiking
right
to
build
a
postsynaptic
or
presynaptic
trace
of
the
activation.
That's
what
we
saw
in
the
voltage
traces
right.
This
is
of
course,
normalized
now
to
some
maximum
firing
rate
also
assumes
to
avoid
under
one
from
the
logarithms
minimum
firing
rate,
but
these
can
be
very
small
Epsilon,
so
you
essentially
get
an
exponentially
weighted
moving
average
filter
of
the
pre
and
postsynaptic
activity.
F
Now,
why
is
that
useful?
Well,
if
you
wanna
do
a
heavy
learning
them
to
learn,
for
example,
this
Co
activation
here
when
assume
is
actually
together.
We
don't
want
to
grow
strong
weight
when
they
are
not
active
together.
So
you
can
do
that
by
doing
this
kind
of
computation
on
the
pre
and
the
post
synaptic
side,
and
you
don't
really
need
to
understand
all
the
different
stages
here.
They
have
different
biological
justifications.
F
But
the
main
point
is
that
you
can
determine
a
product
of
the
pre
and
postsynaptic
activation
traces,
which
is
going
to
give
you
the
heavy
and
component
right.
So
those
are
the
cases
whether
fight
together
and
this
kind
of
product
will
be
useful
for
generating
this
exactly
logarithmic
call
activation
weight,
which
indeed,
then
shoots
up
as
soon
as
these
two
neurons
are
firing
together,
firing
together
and
now
they
are
wiring
together.
This
P
trace
goes
up
which
builds
up
a
higher
beta
similar
for
a
strongly
active
units,
sweat
where
we
have
done
a
strong
prior.
E
F
Strong
posteriors,
in
this
case
you
get
a
corresponding
by
usual
bias
for
your
network.
So
now
we
have
a
way
of
taking
a
spike
train
right
from
a
spiking
neural
network
simulation
and
turning
it
into
a
connective
weights
that
we
can
use.
So
how
do
we
now
turn
this
into
recurrent
neural
network
that
leverages
all
these
awesome
things
about
practice
and
whatnot?
Well,
you
simply
convert
this
theoretical
structure
into
something
big
by
physical.
K
K
F
Despite
the
epsilon
us
and
minimum
assumed
firing
rate
F
max
is
simply
to
normalize
the
Z
trace
to
some
maximum
assume
firing
rate.
The
Tau
here
is
simply
the
synaptic
time
constant.
So
you
can
flip
that
if
you
have
an
ample
synapse,
that
would
be
five
milliseconds,
which
is
lower
signups.
You
know
and.
F
K
K
F
K
K
F
It
gets
very
complicated
because
it
depends
on
the
type
of
coding
right,
because
this
type
of
a
code
obviously
is
similar
to
a
red,
convers
you're
averaging
over
several
spikes
here
and
you
just
come
low
and
up
with
the
kernel.
We
are
doing
that
because
we
argue
as
biologicals
well,
the
the
postsynaptic
membrane
does,
intrinsically
because
of
its
time
constants,
and
what
the
receptors
doing
because
of
their
soon
after
time
houses.
So
these
have
some
biological
justification,
but
they
are
far
abstractions
in
order
to
make
it
possible
to
build
a
network
like
this.
F
So
you
now
have
a
sheet
of
these
mini
columns.
Right,
keep
in
mind
every
one
of
those
nodes
is
again
not
actually
one
unit,
it's
actually
several
spiking
units
and
they're
connected
both
locally
recurrently.
This
connections
particular
layer,
2
3,
but
they
are
also
connected
between
the
different
mini
columns
right
and
these
connections
can
now
be
learned.
So
actually,
what
you're
learning
is
spike
based
connections
here
right,
but
we
now
have
a
site
based
learning
rule,
so
we
can
learn
these
connections
and
that
allows
us
to
build
contractors.
What
do
I
mean
by
that?
F
So
you
know,
I
was
fighting,
you
have
no
network,
you
have
some
cells.
Let's
say
you
know
you
had
all
those
red
mini
columns
mark
right.
So
all
of
these
are
active
with
some.
You
know
random.
You
know
irregular
firing
and
all
of
the
ones
that
on
don't
belong
to
that
are
active
for
some
regular
firing
and
if
we
now
for
stimulate
the
neurons,
you
see
all
the
different
mini
columns
here,
like
16
mini
columns
because
16
hyper
columns
and
everyone
one
selected
mini
column.
F
Then,
over
this
time
window,
where
we
are
stimulating
their
heavy
learning,
you
will
learn
that
the
firing
of
these
units
is
actually
correlated.
It's
going
to
update
the
exponential
weighted
moving
averages,
the
priors
in
the
posterior
and
the
co
activation
for
beliefs
and,
as
a
result,
the
network
now
has
an
imprint
of
this.
This
activation
some
memory
results
and
that
leads
to
our
autogenic
reactivation.
F
So
these
brief
bursts
of
reactivations
that
you
see
here
they
tend
to
actually
become
a
bursting
very
similar
to
what
you
see
in
biology
because
the
basket
cells,
these
strong
inhibitory
tools
that
you
saw
on
the
units
they
will
locally
shut
off
the
activity
as
soon
as
one
mini
card
must
come
strongly
active.
It's
going
to
silence
the
activity
in
its
neighborhood.
F
That
is
not
going
to
stop
the
attractor
after
the
first
time,
because
there's
going
to
be
out
of
phase
excitation
from
more
distant
mini-cons,
which
gonna
deliver
incoming
spikes,
so
there
that's
why
the
attractor
will
keep
going
for
a
while
until
the
transmitter
is
depleted.
So
it's
mostly
synaptic
depression
that
switches
these
off,
and
so
then,
after
some
time
of
recovery,
you
can
get
randomly
activation
events
again
now.
This
is
just
storing
one
happen.
Yes,
so
that's
relatively
boring!
F
F
In
between
the
words,
the
network
can
do
whatever
it
wants,
and
what
you
didn't
see
is
that
patterns
that
you
trained
now
start
reactivating
in
between
the
support,
because
the
network
learns
from
its
own
activity
and
it's
it's
the
never
produces
fine
with
statistics
that
match
the
input
it
has
seen
and
it
we
learns
its
own
earnings.
So
it's
sort
of
an
auto
refreshment
where
my
memories
can
now
last
a
lot
longer,
then
the
time
constants
of
the
learning
will
suggest
because
they
can
self
refresh.
F
If
you
do
that,
much
12
items,
you
can
then
analyze.
Well,
what
do
the
distributions
here,
which
happens
to
actually
recall-
and
it
turns
out
that
you
very
nicely
match
the
free
recall
curves,
that
you
see
in
people
so
typically
there's
a
strong
primacy
and
recency
effect
of
those
twelve
words.
You're
gonna
be
very
good
at
remembering
the
first
three
words
and
the
last
three
words,
and
typically
in
the
words
in
the
middle
of
that
is
fall
out.
F
So,
typically,
when
you
remember
item,
you
are
very
likely
to
also
remember
the
item
just
after
and
just
before,
and
so
these
network
have
a
similar
effect,
and
so
you
can
match
them
up
with
cognitive
studies.
There
is
obviously
a
bit
of
a
toy
application,
but
the
reason
why
this
is
interesting
is
because
you
can
still
match
the
electrophysiology
of
real
recordings
here,
so
you
need
to
answer
other
weights
that
you
are
growing
here
right
with
this
patient
people
are
actually
realistic
might
say.
You're
running
will
now
learn.
You
know
five
million
volts.
F
You
know
postsynaptic
potentials,
that's
you
know
kind
of
unheard
of,
and
so
you
can't
really
have
that
and
we
find
that
in
the
regime
that
these
networks
operating
when
they're.
Finally,
four
five
spikes,
you
get
very
nice.
You
know
one
1.5,
millivolt
excitatory
postsynaptic
potentials,
so
you
can
go
back
and
forth
between
the
ideas
of
cognitive
science,
electrophysiology
and
machine
learning,
and
then
you
can
get
the
published
in
a
very
nice.
You
know
neuroscience
and
capisci
for
these
kinds
of
things.
You
can
also.
K
E
K
F
You've
done
the
sequence:
instead,
that's
actually
what
Phil
Tommy
did
for
his
dissertation.
You
show
that
these
networks
cannot
just
learn.
You
know,
lists
and
show
primacy.
You
either
see
condition
recall,
but
they
can
also
run
for
sequences,
because
you
would
associate
all
these
items
with
one
another
because
we
play
them
so
close
to
each
other,
and
it
is
different
in
my
case,
because
I
actually
intersperse
them
with
science.
But,
like
you
know
what
does
learning
task.
E
F
F
It
is
rather
unreasonable
to
assume
that
the
neocortex
already
has
that
the
structures
that
are
implicated
in
short-term
memory
already
has
structures
set
out
for
everything
you
might
theoretically
learn.
It
is
much
more
reasonable
to
assume
that
these
structures
are
built
online
when
you
need
to
learn
them,
and
so
on
that
principle
you
can
now
use
a
fast
hebbian
approach
to
build
such
a
lot
like
that.
So
now
you
set
out
and
start
okay,
where
do
I
see
stable
on
to
the
memory
of
sorts
that
say
for
visual
items
right.
F
H
E
F
Area
that
is
known
to
not
be
so
specific
and
to
be
fast
changing
in
what
it
is
selective
for
that
dollar
level.
Big
motor
cortex
is
highly
involved
in
working
memory.
Working
memory
tends
to
be
prepared
when
you
envision,
it
may
damage
it.
So
some
speculative
model,
you
could
say,
is
saying:
well,
you
have
some
fast
learning
units
right
with
with
heavy
and
learning
capabilities
now
and
prefrontal
cortex,
and
you
need
to
link
them
with
two
networks
that
have
already
many
many
stored
items,
long
term,
memories
that
are
stable
structures
that
you
can
now
flexibly.
F
You
know,
Co
activate
and
build
it
to
a
pairs
or
whatever
conjunctions
of
item
activations
you
want.
Then
you
take
some
data
on
cortex.
Well,
how
many
connections
are
there
actually
between
those
areas
right?
So
if
you
take
macaque
data,
these
areas
are
some.
You
know
400,
you
know
from
one
another.
So
what's
the
length
of
the
wiring
or
the
connection
delays?
How
many
sign
ups?
Are
there
there's?
F
You
know
good
experimentalist,
you're,
doing
great
work
on
that
and
you
try
to
put
them
together
with
the
normal
sinus
models
of
constraining,
electrophysiology,
which
I
mentioned
and,
all
importantly,
a
heavy
STP
room,
and
then
you
generate
a
big
synthetic.
Moreover,
even
in
something
supercomputer
for
that,
obviously
that's
gonna
be
expensive
to
synthesize
that
much
and
then
put
them
all
on
them.
Some
of
you
who
know
a
little
bit
about
protocol
architecture
will
already
see
that
there's
now
different
cell
layers
here,
super
granular
model
that
actually
differentiates
different
pathways,
and
so
you
are
simulating.
F
The
input
layer
and
you
have
already
embedded
long-term
memories,
recurrent
structures
right
embedded
into
these
in
these
two
long-term
memories,
and
you
now
have
a
flexible,
short-term
memory
system
that
is
connected
with
plastic,
back
projections
and
plastic,
recurrent
connections
to
learn
and
bind
associations.
You
can
actually
use
that
so.
F
I'm,
just
gonna
wrap
this
up.
When
you
do
working
memory
encoding,
you
co
activate
two
items
in
the
long-term
memory
as
a
result
of
that
with
some
small
feedback
delay,
you're
gonna
start
seeing
these
gamma
bursts
happening
in
prefrontal
cortex,
which
we
can
obviously
then
compared
to
macaque
recordings
of
Milla
lab
at
MIT
is
sube
recordings
from
that
one
of
our
PhD
students
went
there
and
then
you
can
ask
well
what
happens
doing
maintenance
right.
F
So
when
you're
not
maintaining
these
items,
does
the
prefrontal
cortex
actually
lead
now
the
posterior
areas
sort
of
to
keep
those
long-term
memories,
their
lives
you're?
Holding
on
to
these
structures
that
you
have
now
found,
and
then
during
we
call
awake
and
see
the
associative
effect
of
these
I
can
just
give
you
a
trigger.
I
can
just
stimulate
one
item
in
one
of
the
networks
say
as
HIV
right
and
then
with
some
feed-forward
delay.
The
prefrontal
cortex
is
going
to
retrieve
that
heavy
and
Laurent
index
to
the
memory
right.
F
So
that's
not
really
the
full
memory
contents
just
enough
neurons
to
bind
the
full
conservative
representation
with
something
back
to
later
prefrontal
code.
It's
going
to
retrieve
their
associated
item
from
the
other
cortical
area
such
that
I.
Don't
need
to
be
able
to
learn
these
connections
here
as
long
as
I
have
a
dynamical
helping
prefrontal
cortex
that
can
find
these
items
and
then
because
the
spiking
networks
provide
a
lot
of
opportunities
to
match
data.
E
F
Models
are
still
evolving
when
they
started
out
as
one
layer
Network
for
classification
they
acquired.
You
know
the
cortical
detail,
we
use
them
for
reinforcement,
learning
to
assimilate
basal,
ganglia
action,
selection,
I
started,
Phil,
told
you
that
I
told
you
about
right,
build
a
spike
based
version
of
the
BPM
learning
rule
and
I've
built
a
lot
of
systems.
On
top
of
that,
that
work
is
not
finished
yet
there's
people
working
with
it,
and
so
let
me
go
tearing.
G
J
F
A
good
point,
first
of
all,
one
should
point
out
that
these
kinds
of
tasks
that
you
were
describing
eclipse
what
working
memory
does
working
memory.
It's
like
three
or
four
items
right
and
then
maybe,
if
you're
watching
them,
you
can
get
to
more
with
that
right.
If
you
leveraging,
you
know
tools
that
you
have
like
the
phonetic
you.thank.
You
know
coddled
egg,
toast
cheese.
You
can
keep
going
away
without.
F
Just
because
you
have
a
dedicated
circuit
for
phonetic
repetition,
so
so
everything
beyond
that
is
clearly
not
working
memory
anymore.
Yeah
and
you're
gonna
need
to
explain
how
that
works.
I
think
it's
still
useful
to
think
about
in
these
associative
terms,
because
when
you
ask
these
people
how
they
do
these
tasks
right,
the
the
most
common
method,
the
method
of
loci.
Think
as
the
palace
of
the
mind
it's
this
idea
of
napping
objects
that
you
want
to
remember
or
sequences.
E
F
Onto
already
existing
structures
which
to
me
is
a
strong
indication
that
working
memory
really
is
about
this
interaction
between
long-term
memory
structures,
we've
already
built
and
you're
leveraging
them
to
put
in
new
things,
and
so
in
some
sense
the
we
abandoned
the
the
original
discussion
of
short-term
and
long-term
memory
in
favor
of
working
memory
in
part
because
of
this,
because
it's
not
really
just
about
the
time
right.
It's
about
the
operability
of
the
memory,
what
you
are
doing
with
it,
and
in
that
sense
it's
helpful
to
understand
working
memory
as
leveraging
the.
F
F
D
E
D
D
F
I
think
it's
one
of
the
biggest
one
of
the
biggest
topics
sort
of
to
talk
about
right
this.
This
wild
order
of
magnitude
on
which
we
have
many
right,
because
we
we
are
encoding
things
like
I,
can
flash
you
a
car
from
a
cart
again
or
nasal
space.
I
put
it
down,
and
you
can
tell
me
what
it
was
even
though
I
just
showed
you
for
200
milliseconds.
F
But
you
will
also
might
remember
things
from
two
decades
ago,
so
you
have
memory
on
the
order
of
six
magnitudes
right,
there's,
no
artificial
memory
system
that
can
do
that.
Yet
somehow
our
brain
makes
it
work
and
it
makes
it
work
by
a
lot
of
filtering
the
vast
majority.
Other
things
that
we
remember
are
forgotten
and
the
synthetic
systems
that
we
are
building
are
not
really
designed
for
that.
F
The
other
person
forget
then
often
catastrophically,
but
this
idea
that
you
should
do
what
these
networks
do,
where
you
have
like
an
exponential
weighted
average
which
dies
down.
So
unless
you
refresh
or
transfer
into
a
more
lasting
form
of
memory,
which
has
a
slower
updating,
you're
gonna,
forget
natural
biology
suggests
that
that
isn't
it
the
case.
So
you
have
prefrontal
cortex
that
shapes
its
activity
and
builds.
F
Selective
settings
like
weekly
mantra
learning,
but
then
you
also
have
two
cameras,
which
can
you
know
a
black
built
change
connections
within
minutes
and
then
you
get
sleep,
dependent,
I'm,
a
conservation
into
a
very
slow
learning
distributed
neocortical
substrate,
and
so
maybe
we
really
do
have
plasticity
on
just
vastly
many
different
orders
or
timescales.
And
if
we,
you
know,
build
synthetic
systems
like
that,
maybe
we
could
get
there
where
we
could
to
have
one
system
too
much
of
learning
but
still
learn
models
of
the
world
to
solve
the
stability.
K
K
F
Right,
so
to
answer
the
last
question:
first
and
no
I've
not
worked
on
that
specifically
I
seem
like
I,
know,
colleagues
who
are
working
on
this
problem
and
it's
very
interesting
to
see
if
we
can
interest
them
in
these
kinds
of
models,
but
I
can't
speak
a
little
bit
further
to
the
first
question.
So
obviously,
yes,
it's
true.
We
are
mapping
somehow
to
elect
discrete
categories
and
well
that
is
fine,
for
you
know
maybe
early
cortex,
like
visual
cortex,
it
gets
full
manic
in
higher
area.
F
What
I
didn't
show
you
this?
How
these
sparse,
activations
and
preflop
Aquatics
actually
look
like
in
terms
of
spiking
activity
when
they
find
items,
because
one
they're
very
sparse,
there's,
like
you
know
four
percent
or
so
of
the
neurons
that
then
you
know
get
selected
and
the
other
one
is
that
they
they
are
very
fast
changing
right.
So,
even
though
it
is
discrete,
we
don't
need
many
more
neurons,
because
in
fact,
working
memory
is
very
little.
Capacity
within
strength
is
not
capacity.
E
F
B
F
So
you
don't
necessarily
need
a
very
big
Network
if
you
have
a
lot
of
flexibility
and,
of
course,
it's
a
little
bit
synthetic
to
say,
there's
long
term
memory,
which
is
eternally
static
and
this
this
super
flexible
memory
system,
obviously
in
the
real
brain,
there's
going
to
be
a
gradient
of
sorts
and
it's
not
really
in
dedicated
long-term
memory
area
and
a
dedicated
short-term
memory
area.
So
all
of
these
are
you
know,
purposeful
simplifications.
F
The
other
thing
with
the
with
the
interval
code
is
a
bit
more
tricky
question
where
you,
how
do
you
do
continuous
fade
variables
if
you
are
now
long
to
represent
a
continuous
distribution
rather
than
than
discrete
months?
In
some
ways,
we've
made
it
very
simple
for
us
and
that
we've
leveraged
the
powers
of
attractors
here
pattern
completion,
and
so
we
have
discrete
object
and
neither
the
object
does
get
recalled
because
the
attractor
kicks
off
or
does
not
get
recalled,
because
the
tractor
does
not
kick
off.
F
What
do
we
do
with
the
places
in
between
it's
a
fascinating
question?
I,
don't
think
this
I
don't
have
a
clear
answer,
but
at
least
not
on
the
basis
of
an
attractive
theory
of
neocortex.
Obviously
one
can
argue
about
the
tractor
models
that
you
know
have
nicely
shifting
representations
I
find
them
somewhat
artificial.
F
The
thing
is
also
the
whole
thing
about
neocortical
architecture
is
is
very
discrete
right.
All
of
the
neurons
inside
a
mini
column
are
not
just
selected
for
the
same
item.
They
have
the
same
genetic
progenitor
cells,
so
they
are
built
as
one
unit
and
so
I
think
we.
We
cannot
call
out
by
these
other
tractor
models.
We
we
have
to
say
what
these
units
are
actually
doing.
No
Manta
has
a
much
bigger
visual
for
that
in
part
because
they
they
attribute
a
lot
more
power
to
the
eventual
individual
neuron.
F
B
K
F
F
You
immediately
see
the
deficits
when
these
people
have
any
amount
of
minute
catalog
damage.
So
to
me,
that's
a
strong
indication
that
you
can't
divorce
a
strict
definition
of
working
memory
from,
but
of
course,
we
are
always
working
on
that
edge
right.
Part
of
what
you
know
you're
seen
here
today
will
be
encoded
in
hippocampus.
You
know
some
of
it
is
gonna
linger
around
a
little
bit
in
terms
of
reverberating
activity
is
some
of
it.
You
know
laughing
synaptic
changes
and
maybe
some
small
fraction
that
is
hopefully
gonna
come
solid
internet.
F
I
mean
this
is
a
story
that
you're
in
deep
sleep
when
you
look
at
pattern,
reactivations
is
very
nice
intractable
in
case
of
play
cell
experiments,
you
get
hundreds
of
repetitions
of
prior
taylor
and
episodes,
then
the
more
important
some
particular
episode
the
morning.
Kids
we
play
and
it's
like
hundreds
and
thousands
of
repetitions
so
like
these
are
females
events.
F
So
there's
like
three
four
of
these
sharper
footballs
every
second
and
the
whole
time,
while
you're
tipsy,
so
that
that
is
a
lot
of
replaying
and
about
that
you
could
never
practice
that
much
in
part
because
of
compresses
whole
paper
episodes
into
these
100
millisecond
long
nested
oscillations.
So
we
have
very
strong
evidence
that
it
becomes
doesn't
heed
to
that
consultation.
It
is
also
clear
that
you
don't
necessarily
need
it
for
working
memory,
but
we
somehow
me
to
explain
the
bridge
where
we
violate
the
working
that
we
can
take.
F
G
G
And
useful
Mira
real
from
our
desires,
invented
to
accelerate
the
doctrinal
Geoscience
findings
and
principles
in
efficient
engines.
We
believe
that's
what
it
takes
to
take
to
get
to
our
toys
and
presents,
but
we
want
to
promote
a
neuroscience
inefficient
engines
no
times.
So
we
don't
talk
only
about
the.
G
The
brain
and
then
we
move
on
to
mission
learning.
So
if
you
are
here,
it
is
a
machine
learning
researcher
and
flier.
That
was
a
big
heart.
If
you
really
love
it,
it's
been
fascinating.
Like
I
did
this,
we
have.
The
thing
is
they're
here
every
day:
I'm
learning,
pigmented
and
I
find
it
fascinating.
How
much
we
talk
about
the
pink
with.
G
If
I
can
so
Thomas
be
funny,
sorry
fact
right.
It's
a
research
sites
like
google
apps
and
it's
also
affiliated
with
the
Neuroscience
Institute
and
in
the
past,
was
a
research
fellow
at
Harvard,
Medical
School
and
at
the
French
National
Center
for
scientific
research.
Here
in
his
PhD
from
University
of
Birmingham,
we
saw
actual
an
evolutionary
awkwardness,
robotics
and
artificial
life,
the
most
prolific
research
of
dozens
of
papers
published
in
Maine
citations,
and
we.
L
L
L
L
L
That's
the
first
time
system
is
actually
exploring
the
base
it
finds
and
then
immediately
it
goes
back
to
that
cache
and
then,
after
a
few
minutes,
because
that
is
very
fast,
it
has
memorized
that
dress
the
configuration
of
the
Nez
Perce
or
the
path
to
the
food
cache,
like
maybe
two
three
tryouts
physically
the
game
is.
It
has
never
seen
before
those
thousands
of
examples,
know
millions
of
tryouts.
E
L
L
Nevertheless,
and
in
our
platypus
pretty
sounds
if
you
have
two
sets
which
tend
to
be
prevented,
yes,
should
constantly
grow.
Whether
is
there
tend
to
be
consecrated.
The
content?
Decrees.
Now
something
that
strikes
the
tree
striking
on
this
type
of
fruit
is
that
there
are
two
transfer
that
is
just
local
ferry,
nothing
else,
and
yet
this
very
basic
unsupervised
learning
at
the
doral
state
results
at
the
scale
of
the
ordinal
brain
into
very
efficient
goal,
directed
beyond
behaviorally,
effective
learning.
So
how
does
it
present
possible?
L
E
L
Has
been
designed
across
it
very,
very,
very
good
candidates
by
evolution
in
order
to
present
about
structure
in
which
this
specify
starting
recent
instances
are
useful
and
efficient.
Whatever
she
has
done
in
general
is
pretty
much
everything.
Of
course.
In
particular,
it
has
designed
the
overall
community
pattern
between
the
areas
of
the
brain.
It
has
set
the
initial
weights
for
the
connections
and
this
mystical
sense.
It
has
also
set
different
amounts
of
plasticity
in
different
areas
of
the
brain.
The
cortex
is
very
fast.
People
confess
in
front
of
more
cortex
is
very
story
to
us.
L
If
you
put
the
time
to
lots
of
different
types,
yes,
dependence
and
the
brain,
and
because
one
of
the
season
carefully,
designed
and
soon
by
revolutions
revolution
across
a
very
imperative
the
time
this
was
a
simple,
very,
very
powerful
and
efficient
learning,
coordinator
learning
and
the
scale.
So
what
we
want
to
do
here
is
take
inspiration
from
this
when,
instead
of
using
evolution,
the
design
on
it
was,
we
want
to
use
exactly
the
same
factory,
so
beautiful
results.
L
L
So
this
is
a
quick
how
many
ways
to
do
it?
Yes,
the
way
we
came
up
with
this
is
the
stop.
The
normal
equation,
for
the
exhibition
from
your
own
is
just
under
a
normal
equation
to
offend
any
neural
network.
The
output
of
a
neuron
is
equal
to
the
weighted
sum
of
its
inputs
passed
through
some
kind
of
minority.
L
Exactly
the
network,
the
only
thing
that's
different
is
what
we
put
in
the
weights
detection.
So
normally
you
start
up
your
network.
You
only
have
this
w
here,
each
condition,
there's
W
bar
itself,
which
is
trained
by
reading
this
and
tells
you
how
strong
this
collection
is
once
the
tourist
trains,
that's
it
between
two
fixed
individuals.
Here
we
still
have
that
we
still
have
a
baseline
so
for
birth
weights
attraction.
We
don't
talk
about
that.
We
have
a
life
time
varying
quantity.
L
D
L
L
L
This
may
be
a
bit
fuzzy
I
just
wanted
to
show
that
the
whole
thing
is
very
simple:
first
effective
of
the
algorithm
for
training
works,
yeah
I
mean
here
I'm
using
episodes
and
lifetime
interpreter.
It's
the
same
thing
when
you
show
examples
of
what
economies
are
basically
four
across
lifetimes
and
the
beginning
of
each
left
amount
is.
Rather,
you
said,
although
the
hidden
traces
to
0
and
then
during
the
episode,
you
simply
run
your
network
at
each
time
step.
If
you
get
a
reward,
you
store
it
or
if
you
have
a
certain
grocery
store
yourself.
L
You
date
the
hidden
trace,
which
means
to
change,
keep
doing
that
for
what
reason,
at
the
end
of
a
certain
reward
or
loss
that
you
are
actually
related
during
this
episode,
and
you
take
the
gradient
of
that
across
alpha
and
W.
This
is
again
the
initial
weight
and
the
amount
of
testing
in
extra
function,
and
you
take
a
trillion
step,
and
you
start
again
in
the
metaphor
of
evolution
and
life
from
running
this
year
of
your
own
evolution
across
many
lifetimes,
and
this
is
a
vector.
L
Yes
and
begin
to
show
this
really
really
great,
simple
and
showing
actual
implementation
on
the
stand
on
the
rig
from
your
networking
by
Josh,
for
lens
of
by
George
does
only
need
to
make
plasticity
in
the
regulatory.
All
the
rest
is
torn
up.
In
addition
for
your
network,
the
only
thing
that
you
need
to
add
is
a
flat
confidence,
w-wait
you're
such
need
to
ensure
the
Hadean
trace
of
each
connection.
Then
you
take
into
account
hidden
components
for
the
activations
and
your
dependent
race.
That's
it
for
interpretation
all
the
time.
L
L
Two
images:
three
meters,
you
need
to
memorize
them
and
then
I
show
you
proportional
version
of
one
of
those
emails
and
you're
going
to
reconstruct
the
missing
from
the
morgue
again.
I
insist
on
the
fact
that
we're
not
just
training
is
singularly
Network
to
memorize
a
fixed
set
of
images.
It
was
I
admit
that
when
they
are
seen
during
training,
we
will
want
to
train
to
memorize
arbitrary
nobody
messes
with
them,
and
we
show
that
this
that
what
is
it
just
says?
L
L
L
L
Well,
so
we
verified
that
being
able
to
actually
train
diversity
of
information.
Is
it
reducible
to
verify
components
that
I
showed
you?
We
will
do
some
image
being
able
to
Train
individuals
of
each
connection.
Makes
you
better
at
this
and
that's
what
we
should
experience.
It
can
also
make
you
much
better
all
the
tasks,
all
right.
Yes,
this
is
the
other
task.
L
That's
it
or
so
the
dream
desk
was
missing.
Creation,
I
have
civilized
life
from
your
name,
assure
you
animation,
and
once
you
construct
its
the
intro
for
reinforcement,
so
we
trained
a
very
simple
task,
which
is
the
main
expression
desk
we're
seeing
about.
We
literally
showed
you
first,
so
it's
an
amazing
correctional
desk.
L
The
major
score
is
this
a
very
simple
implementation
from
one
person
to
the
next,
when
we
have
a
certain
reward
which
can
be
in
any
location
with
the
agent,
which
is
this
yellow
thing
here,
its
goal
is
to
find
the
reward,
and
once
it
has
found
this,
it
gets
teleported
at
a
random
location.
In
amazement,
it
needs
to
go
back
to
everyone.
Teleporter
beginning
is
to
go
back
itself,
so
the
maze
is
always
the
same.
So
during
your
life,
you
don't
need
to
memorize
the
maze
it
can
story
so
to
speak
in
your
own
innate.
L
But
the
position
of
the
what's
been
Trinity
moment:
the
only
thing
that
this
engine
can
see
it's
a
small
neighborhoods
three
times
three
around
itself,
so
at
first
ones
ducked
into
the
bed.
It
doesn't
know
whether
we
want
is
means
to
explore
five
debates
find
out.
Why,
then
I
guess
important
go
back
there
as
often
as
possible?
Your
name
is
our
main.
You
show.
You
of
course,
is
completely
on
the
other.
L
E
L
Okay,
so
now
we
are
shamans
plastic
and
first
constraint,
and
no
arbitrary
printable
seem
like
they're
not
seen
during
that
time
very
well,
but
in
the
bank
plasticity
it's
Spencer
device,
but
it's
still
controlled
by
the
rest
of
the
virus,
chemicals
in
the
brain
which
determine
at
each
point
in
time
then
to
the
same
directory.
That
was
a
magnet
universal
sign
of
the
plasticity.
So
we
have
this
fantastic
4,
which
determines
roughly
how
much
change
should
be
applied
to
certain
condition.
The
because
of
your
delayed
us.
L
The
brain
can
control
at
any
point
in
time
when
to
be
blasted.
They
see
if
you
want,
we
all
know
it
sometimes
really
some
things
which
are
very
important
and
ignoring
or
the
things
which
are
less
important.
Those
are
pieces
of
the
interconnection.
Your
silence,
the
simplest
version.
Is
this
one?
Well,
this
is
a
process,
the
experiments
when
you
basically
stimulated
some
connections,
and
you
show
that
there
is
an
increase
in
the
20
inch
weights
between
between
these
two
cells
without
mean.
L
E
L
Yes,
there
is
in
the
network,
yes,
one
of
neurons,
which
are
the
signs
to
be
the
computers
of,
inter
how
many
ways
you
can
do
it.
You
can
a
single
interval
for
Network
different
possible
network
and
different
investors
to
do.
But
the
important
thing
is
just
many
bitter
is
the
rate
of
Christian
one
can
control
its
own
plasticity
in
retail.
L
L
So
in
conclusion,
we
show
that
with
gradient
descent,
we
can
train
networks.
It
can
keep
longing
after
the
initial
training,
complete
diversity
that
has
never
seen
during
the
training
and
with
automatically
and
efficiently
and
we're
sure
that
they
can
acquire
water
for
the
more
passive
networks,
including
in
decisions
always
desk.
So
we
show
that
this
is
efficient.
L
K
L
L
This
is
from
the
dopamine,
was
like
before
not
before
humans,
nothing
changes,
but
it
was
applied
long
after
the
events
has
no
change
when
the
dopamine
is
eveni
just
one
second,
after
the
events
of
any
solution,
gotta
be
implemented.
That's
it
isn't
original
paper
that
basically
shows
that
it's
very
easy
to
pronounce
this
thing
is
very
flexible.
See
you
you
just
especially
equation,
can
literally
write
by
talk
and
just
works,
so
yeah,
there's
other
things.
That's
very
interesting,
trying
to
find
more
flexible
and
more
complex
forms
of
vegetation
and
see
what
the
SIS
can
doing.
L
F
E
F
And
then
maybe
one
more
question
about
you,
you
use
the
clip
function
to
avoid
the
run
away
off
your
heavier
weights.
Obviously
there's
different
ways
to
get
around
them
and
like
I,
get
around
that
the
logarithmic
weight
so
that
function
into
this
guy
and
obviously
in
in,
but
there's
many
different
approaches
to
this.
L
Values,
homes
are
so
sleeping
where
was
always
increased
within
Christmas,
smaller,
smaller
ones,
a
circuit,
that's
various
ways
how
it
was
just
as
well
of
its
own
competitiveness
with
we
have,
but
they
say
that
the
reason
why
it
works
so
well
is
because
we're
not
just
throwing
your
fatigue
in
the
system.
We
ourselves
SGD
greater
distance
will
force
the
network
to
remain
in
the
regime
where
keeping
another
problem,
because
against
the
whole
mentality,
because
you
can
dismiss
I,
don't
also
loop
or
sale,
one
can
be
a
bit
more
tolerant
with.
L
Improvement,
it's
a
big
improvement
for
most
us.
It's
not
miss
it.
Those
are
standard,
neural
networks.
We
are
as
crystal
patients.
As
you
know,
recently,
there's
been
a
lot
of
more
complex
architectures
attention
based
mechanisms,
trust
immerses
itself,
it's
not
quite
chaotic
elasticity,
so
we
also
have
be
improvement
of
those
big
new
things.
I,
don't
know
four
separate
from
networks.
If
you
requested.
I
A
part
of
the
way
that
continues
to
adapt
to
that's
what
the
plasticity
is.
Is
there
any
theoretical
justification
why
you
wouldn't
sort
of
continue
to
dial
it,
the
more
toward
continual
learning
and
less
toward
that
initial
learning?
Like
that,
a
certain
point,
you
could
just
have
it
all
of
the
wavy
plastic
and
have
it
be
just
that
with
in
the
beginning,
because
it
seems
like
yeah
the
first
part
of
the
week.
That's
learned
during
the
training
session.
I
L
E
E
I
I
E
F
Really
a
thing
that
is
also
sculpted
out
of
the
much
larger
Network
right,
like
the
connectivity
of
a
three.
Your
child
is
roughly
a
hundred
times
of
what
an
adult
has,
and
that
is
the
result
of
a
massive
explosion
of
connectivity
post
birth.
So
in
that
sense,
sort
of
the
the
substrate
that
we
have
to
work
with
to
learn
on
a
daily
basis.
Now
that
we
are
adults
right,
it's
actually
a
network
that
has
been
sculpted
out
of
a
much
more
complete
network,
and
so
in
that
sense,
maybe
it
would
be
worthwhile.