Thanks it's really been
a pleasure to be here.

Told several of you have a sixteen year
old who is interested in engineering

programs and we didn't tour
engineering schools and we looked up

jeering program on the Web you know
trying a tech always came up on top.

Is in place the next sun with is it
Georgia Tech we didn't make it quite that

or so but
it's it's really an impressive place you

guys are lucky to to really be
able to call this your home.

Of a little bit I saw this
is really impressive so.

All right so today I'd like to tell you
about some relatively recent who are in

the laboratory and maybe it's a little
bit too dark my it's lights are not.

OK good right and so.

In the laboratory in really are our
work is focused on mammalian

North circuits focus on the cortex you
know and as you as many of you know

cortical circuits are part of cortex
this parcel aided in into multiple.

Functional regions.

And.

If excuse me one more request maybe you
could just turn off the lights behind this

lights just to enhance a contrast if
possible wonderful that's good so

cortex is postulated interests
distinct distinct functions and

if you kind of drill into cortical
circuits you have complexity cell types

that are connected in highly
specific circuits and

then of course you have complexity
also on a measure of scale cortical

regions are connected to each other
to form a regional circuits and

as you know action potentials represent
information in the brain sambal so.

Neurons that fire action potentials
represent information in the brain and so

we'd like to really understand in the
laboratories how these will solve itself

neurons action potentials and so
most neurons really represent information

what drives these representations
of information on the one hand and

then how does the animal actually
use these representations for

behavior would you dress most of
these questions in behaving mice and

more specifically we study
circuit computations in this so

much a sensory cortex and like.

As you heard from Stanley
it's laboratory work for

example meaning in the context of
risk of a so much a sensation and.

So much a sensory cortex it's really all
about fast times you think milli second

time computation essentially signal
processing from the sensory for

periphery and over last five years and
this would be really the this would be

the focus of my talk we've really
also looked at frontal cortex area.

In particular an area called.

Implemented slow times it's related
to loosely speaking deliberation and

short term memory motor planning things
that happen all the times of second and

one of the things that
I'm really interested in.

Something that we'd like to figure out
over the next five years how with very

similar kind of harks where some of kinds
of neurons some of kind of by physics do

you get actually slow times in these front
of cortical areas and these very fast

times it's in sensory areas and that's
something that at the very end I'll show

you some stuff hot off the press that
might have some bearing on that OK So

we talk really mostly about motor
planning short term memory today in front

of cortical areas so let me start by
explaining what motor planning is and

why you should care body is so
as as you know many movements are.

To wrap it the right to really be amenable

to online correct send these movements
a cycle physically experiments have shown

in some sense preprogramed and really
evidence for such a planning a motor

planning process comes from behavioral
experiments so when you give

a subject time to plan movements of
subsequent movements for example baseball.

Faster and more accurate and
people have shown this

this goes back to the seventy's and
eighty's mainly human cycle physics and

it's also been done in non-human
primates to some extent.

So motor planning is therefore an
important aspect of motor control any can

perceive the movement by many seconds
right for example during the planning and

right up to a baseball pitch now second

equally relevant is that more planning
is also term memory you could think of

it as a prospective short term memory that
links past events and future movements so

in this picture you have
a picture that plants

a picture instructed by a century sensory
information from the capture the hand

signals catchers and the hand signals of
The Catcher in many settings of course.

Elapsed between instruction.

In the instruction motor plans that
presumably stored over this time skids of

seconds in the pitcher's brain somewhere
nor representations in this brain

bridging the time from the instruction
when the hand signals to the movement hear

the pitch or Right now let's
go to neurobiology right so

what is really more in
the context of the brain well in

the law up has been studied in
the context of so-called delayed

response tasks OK tacit which is sensory
stimulus instructs an action but

to the action is only executed
after some memory peer.

In the technical jargon of field
delay Apoc is strong like you see

here this is a really famous experiment
is really the one of the first cognitive

neuroscience experiment actually one of
my favorite experiments in not just this

you've been in earth science in general
does from Everett from the seventy's and

what they did is they instructed a monkey
to either push or pull a lever but

the monkey was supposed to execute the
push pull movement only after a delay of

popular six seconds and
you can see something curious here and

you can see perhaps the activity that is
now referred to a sprat Atari activity

which is in thought to be the neural
correlate of motor planning this non

seems to intice a pate the push movement
instructed to push push this right here

this neuron action potentials or
spikes Spike rate is elevated

long before the movement anticipating
when this they're on actually shuts off.

During the movement OK so nor on that
anticipates that particular kind of woman

this is specific to the movement because
when the monkeys instructed to pull and

it he pulls this on is
actually Silence OK so

it is a pain specific neuron movement and

it's movements that the non cares about
rather than instructions because when

the monkeys instructed to pull
about he pushes spiders elevated so

it's related to the Future Movement OK to
you rather than the past and structure.

To predatory activity this normal carload
of Mordor right is really activity

that anticipates future movements often
seconds or longer before the movement.

Now in the brain as you know
short term memory syndrome

are represented by changes in Spike rates
that can be maintained in the absence of

sustained input over time skewed so.

Seconds and preparatory activity.

Is an example of such a memory
craze right here is instruction

preparatory activities maintain this is
an example of a cellular memory trace and

so for these reasons
the mechanisms underlying motive

in preparatory activity I think of a broad
significance in brain research in F.

received a lot of attention
over the last forty years.

So the kinds of questions that we want
to ask in rodents what we can do more

comprehensive Fanaa this is sort of the
phenomenon more making this economics is

how is predatory activity organized
spatially and temporally is preparatory

activity this stuff is actually causally
related to future movements right or

is it just at the phenomenon and perhaps
most important for the purpose of my

talk you know how do North circuits
maintain preparatory activity and

here the mystery is that neurons
have time constant so many seconds

this century memory less but the circuits
here maintain memory of the stimulus for

Motor plant over time skids of seconds
tens of seconds this is an emergent

phenomenon of cortical circuits How
does that happen OK And by and but study

motor planning in mice we can begin we can
begin to address these kinds of questions.

So we started by developing
a delayed response to.

Mice this is almost identical to
the kinds of delayed response stars

that have been studied in non-human
primates for twenty thirty years or

so here my it's perform a sensory
discrimination an auditory attack

the discrimination I want to talk about
decision making all that kind of stuff

that's interesting it's own right but the
signal of their decision about the sensory

system is by either leaking left or
right to a target OK but

they only make that decision known by
looking left or right to a target.

To delay a part that is imposed by the
experiment that can last seconds this task

is now pretty widely used in the field in
this odd looks in a recording a part or so

here you have a model performing the tasks
this is actually tactile version of

the task and the details so with sensory
discrimination that don't really matter

the pole comes within reach and one of the
mice touch the pole with his whiskers and

then dutifully sample the mouse holds
its time before signaling its decision

with scared direction licking and these
are individual licks you these are trial

it's arrayed in the vertical dimension
time in the horizontal dimension so

as behavior output OK we measure
this skewed direction leaking and

throughout the top Blue Ridge first right
movements red to leftward movement OK.

And behaviorally experiments like
in previous experiments from

non-human primates show that mice
actually use the delay Apoc to.

Plan precise and

accurate movements is just wrong in one
experiment here that if we increase.

The mouse time to plan movements by
lengthening the layabout the subsequent

movements become faster the reaction
time becomes shorter and

also the movements become more accurate so
this task is robust and

quantitative it really provides a tool to
dissect the neural mechanisms underlying

motor planning short term memory knowledge
in my eyes All right let's go to neurons

now in mice OK So as in the tangent
Everett's experiments memory related

neural signals is now nor on record
from the frontal part of the cortex

called motor cortex I'll tell you
a lot more about that that shows

preparatory bonafide preparatory
activity so here now you have a right.

I think all of human have seen Ross
applauds you saw no trial separate

in the vertical time in the horizontal
dimension each circuit you corresponds to

an actual potential OK right for
a trial left for trial does this when

the on a move this is the memory business
when the instruction decision making so

to speak or albums and these are the spike
rate that's corresponds to these rosters

and I'll show you a lot about all of these
spike rates and very relatively few of

these rosters OK so the spike creates a
extracted a simply just the rate of action

potentials averaged across tried so if you
put the purposes of the stock you can see

that this neuron starts to
supporting right become selected for

right movements maintains
the selectivity of park and

shuts off in per movement period
importantly as required for

preparatory activity on trial it's when
the elements instructed to move left but

actually moves right the neuron becomes
against selected So this is nor on that

it dissipates specific movements rather
than caring about the past and structure.

So this stoats us really that we've
developed a mouse system to study motor

planning and short term memory so
we're Can we take this as

more planning and
Predator have been studied for

nearly forty years in non-human primates
and I think what our studies in mice have

provided is a complementary set of
mechanistic insights and I give you some

of the highlights of the last we've been
working on this what about for five years.

For us MI five here is to give
you some of the highlights

some publish stuff interspersed
with unpublished stuff

together with some things that are very
much in very much early stages.

OK so we started with local
localizing preparatory activity and

motor planning and
I think this is really very.

Important because I think localization
is absolutely critical of both

behavior essentially kind of a computation
and a particular activity pattern for

mechanistic analysis of these things you
need to know where it is in the brain or

how widely distributed OK And so
we look like it's more of planning and

predatory to do a small part of
the motor cortex refer to a.

Motor cortex who can so how do we do
this the brain is a big place and so

what we started off is by doing
a loss of function experiment so

the mouse brain accommodates
about one hundred.

Million neurons we think
on the order of sixty or so

brain areas we wanted to ask which
brain regions are critical for

planning and movement OK so
which brain regions the critic.

For the kind of behavior that I've shown
you it with up to genetics you can now do

this in a spatially and
temporary precise manner and

just one slight about math because
this math that we use over and

over again OK so
it's going to be important.

It's something called inhibition
it's our way to inactivate small

chunks of cortex we put channel
drops in into Governor Jack neurons

transgenic mice is relatively
straightforward some of you know this very

well then we can shine light onto
small chunks of tissue like so

Dr Lawrence coverage of
neurons all our local OK and

it's really the excited to the neurons
that carry information out of the trunk of

tissue excited to learn so
silenced when these neurons inhibitor and

so driven so
we can with time scales of tens of many

seconds with time precisions of tens
of thousands of missing this take away.

Millimeter skin chunks of tissue
corresponding to individual brain areas

in a highly reprehensible and
reversible manner so

the activation method said and
it turns out that these transient local

perturbations also very informative and a
powerful tool to probe circuit mechanisms

which shall become important data so
we do this in a comprehensive way we.

Activate one of fifty five
cortical regions spanning most

of the usual spot suspects that might be
involved in motor planning including part

of all parts of the motor cortex much the
sensory cortex and the parietal cortex and

other brain regions one braid
brain region at the time and

then look for when we silence that
chunk of cortex during the delay.

Which silence in which brain
region might have an effect on

future movements during
the Perry movement period and

it turns out that the experiment
the result is relatively simple

it turns out that in activating here
the magenta region in this plot.

During the delay and subsequently causes
an see versus bias in movements and

silencing any other cortical
region is singly on combination

has no effect and this really is
a functional definition of error I'm.

Sorry I'm supposed to stand here and
point this of biology sort of there.

The this is essentially
a functional definition of.

Motor cortex as the brain region
at least in this task critical for

motor pry into so now how does this now

relate to prep atory activity
the neuro carload of multiplying so

we perform an experiment where
we map neurotic to the T.

now in.

The motor cortex and

also other regions with electrophysiology
an imaging to ask how widely is.

Quickly.

Should probably.

Kill the wife.

So how widely distributed is preparatory
activity how does it compare

to the functional definition of in
the context of motor planning so

we can do now with new.

Genetically encoded Propes you can go and
you microscopes you can do large scale

imaging you can image a large
numbers of neurons simultaneously

extract it to the patterns corresponding
to individual neurons and then correlate

the sixty patterns to behavior for
example you can do similar things with X.

a solo electrodes doesn't
make as nice of a movie and

then give you just one piece
of data from the imaging and

the electrophysiology is very similar So
here we took that kind of data and.

Plot it in terms of what the neurons
Skoal to the not they're selected for

movements of selected for
the sensory stimulus and

how they're spatially distributed
here across the motor cortex and

this is time in the time in the trial

movement on set is at zero and
you can see this perception decision

making this flow of information from
sensory information from the more.

Pulse Syria areas to more.

Area and then the emergence of movement
selective activity here in this.

Area mainly in the deep players I'm
going to stop this right here so

the earliest onset of movement related
activity before the movement that is

preparatory activity is really limited to
a very similar region that overlaps the.

Motor cortex so this is parsimonious
decision of convergence year of

being a place where preparatory activities
Express and also the region were lots

of functioning experiments perturb future
movements and I think identifying is

already mentioned the key brain regions is
a critical step in mechanistic analysis so

now let's perform a detailed
analysis of preparatory to just show

you one way we think about
preparatory activity there so

on the left is the neuron that
you've seen earlier on the right for

additional neurons Paris to miss
time histograms of former answers to

highlight kind of the diversity of stuff
that you might see on one electrode in

tuning the neurons so they're nonce that
are selected for right work movements or

left and there they come in about
equal proportions on each side so

that's why here in this schematic
we've got blue triangles red triangles

corresponding Lefferts selective
neurons and they're kind of salt and

pepper here in the Moreover
you have neurons that

are selective early that become
selective hearing the sound.

CALL instruction book when the sensory
stimulus is there are others that maintain

selectivity during the delay of clock yet
others that seem to become selective.

Yet are the neurons that
are most selective in

the Perry movement period perhaps
consistent with driving even a more.

Street motor centers now of course my
ace used to hold an oral population for

behavior not individual neurons right so
how do we make sense of this complexity so

and it turns out the dimensionality
reduction methods really provide

a remarkably simple picture of predator
activity especially in the park and

the relationship to behavior in
particular during the delay at most

neurons either ramp down through
movement or ramp up to a movement and

we will focus now on
memory related activity

popular record typically thirty nine
Ronson Edom summiting is to be with.

Silicon electrode this is now X.

or so a recording so OK for
most of what I say from now on is X.

or so recordings imaging won't
make another appearance and

you can think of a tit T.

of thirty neurons essentially
corresponding to a thirty dimensional

space where each dimension in this
space corresponds to the firing rate of

one neuron and

then during behavior activity patterns
correspond to project three So in this.

Space and the trajectories for right
movements and left movements are different

on average because different neurons
at different firing patterns are rich

in this activity space and
so we can define a coding

direction as the direction that
best separates dynamics in the.

Space for right and left.

Referred to this is the coding direction
it's a simple discriminant analysis for

the aficionados here and
then we can project the activity

to this coding direction that's
shown right here again in this

projection this is by construction
the trajectory separate nicely

during the delay at pop
up to a movement OK And

more importantly activity in this
direction actually accounts for.

Nearly fifty percent of the variance and
it turns out it's all of the variance

that is relevant at least to
the direction of movement and

to the timing of movement we think so

the activity along the coding
direction explains behavior OK Let me

show that to you with an analysis
of seeing the trial it's OK so

here in the schematic now we have
these average activity trajectories

in the coding direction OK up to and
off the delay on movement

onset now the mouse of course has
to use in the trials for behavior.

And if if there is a relationship between.

Activity these activities trajectories
in this coding direction and

behavior we expect the relationship
between single try to project trees and

behavior and the symmetry trajectories
vary from truck to trial so for

example if the trajectory location close
to the onset of movement mightily for

the direction of movement so here is
a trajectory that's close to the average

right trajectory a little bit beyond what
we might expect that this corresponds to

rightward movements with high probability
footer objects we eat towards the left

projects he's for right trajectories
that move towards left trajectories or

closer to left or directories maybe
we see left with some probability and

so on and so forth and
conversely for left trajectory and

this is exactly what we see in the data so

if the trajectories in the civil trial are
closer to the average right trajectory we

get correct behavior movements with
high probability conversely for

leftward trots and as the trajectory
C.E.Q. to the trajectories for

the other movements errors creep in so
that's pretty good evidence

that a lot of activity
actually is related to future

movement that this is really causally
related to future moon this is really

a key area for motor planning but
there is still a.

Some possibilities that this area where
we did this analysis might inherit

these fluctuations from somewhere
else might just sort of and

it might not actually play causally.

Role in movement so
we thought why don't we just introduce

perturbations and
fluctuations directly into the network and

then do the same kind of trial
by trying to novices and

we've done this experiment about half
a dozen ways and the take home message is

always the same just give you one
experiment which I think is particularly.

Illuminating here we tragically
inactivate in such a manner that

we completely obliterate preparatory
activity on average Ok so

the trajectories on single trial
is just illustrated if some.

For some right trajectories on
a memory of what the trouble was and

now meander randomly subsequent
book in the coding direction and

might equally likely end up here closer
to the rightward movement trajectory or

the leftward movement trajectory yet on
a trial by trial manner we can still read

our movement direction with
pretty good reliability.

So by the by analyzing the perturbations
of the effects of perturbation both

the normal dynamics and behavior we can
show that I think preparatory activity.

Instructs future movements.

All right so let's now go to a more
mechanistic analysis of all maintenance

of preparatory activities so
I was present for activity generated and

maintained in actual neural
circuits all of the answers so

I'll just give you sort of an update
of where we are what we've done and

how we're thinking about the problem so
we.

Have discovered really by
accident mechanisms underlying

robustness in maintenance or preparatory
activity and that's that's what I'd like

to show you first and so we discovered
that after large scale in activation of or

activation of a Lem and I mean we
can really silence one hemisphere of

a limb transiently that preparatory
activity or active it with China activity.

In that activity and selectivity recover
in detail over time scoots of hundreds

of millions seconds and that's interesting
experiment seems a little crude

it's an interesting experiment because it
really provides for new challenges for

circuit moderates of predatory
activity and short term memory so

let me show you how this looks
in detail so here is one trusty

neuron here again neuron That's right
what selective surrounds up to it's

persistently active over time
scales of a second or more and

here is the inactivation experiment really
silence it and all of its neighbors.

Entire on one side
activity levels go to zero

as they should because of
the silence after offset of silence

activity recovers over time skids of
hundred milliseconds advertised but

not just activity recovers but selective
if he recovers right on right for

it movement the neuron becomes as active
as it would have been in the absence of

birth to be in the same on leftward
movements some of these neuron and

its neighbors maintain a memory of what
its activity should be despite the massive

disruption of network to and this
behavior holds at the level populations

is again the kind of dimensionality
ruction Earlier know.

That someone is recorder
population along the.

Morning direction the hear perturbation
the selectivity completely collapses but

then recovers over time scales of hundred
many seconds and with this kind of her to

be also the behavior recovers so
how does that work well it.

It turns out not surprisingly that
the memory has to sit somewhere outs

right another area has to communicate
with the perturbed area and

help the perturbed air recover and
we did a search for these areas and

it turns out it's the other
side off in that So here is

on one side and I'm on the other side
in the two sites of the motor cortex

eleven particular incredibly strongly
connected through the corpus close and

there's an anatomy experiment a trace
it was injected right here in the set

the axons project to the other side
of the brain and a whole kind of

whole model of those regions on the two
sides are strongly interconnected so

this is a candidate area that might
help maintain motor preparation in the.

During the perturbations and
it turns out as I've shown you earlier

that preparatory activity for rightward
and leftward movements is on one side and

it's also on the other side because of
this is so when you perturb this side

it could be that effective if he
recovers through the other side and

that's in fact what happens if we know.

Both sides of the selectivity.

So the key principles that underlie
robustness of prefatory activity

are really redundancy and
modularity OK that's

perhaps the more important point here the
circuit consists of at least two more you

see in this case the left molecule in the
left hemisphere in the right hemisphere.

So that when one molecule
is perturbed the other.

The perturbed.

The recover and we refer to these
as module it's because each module

can maintain preparatory
active it independently

when the perturbed when the other
module is perturbed OK so.

Because each more you can maintain
preparatory activity in the pen to mean

redundancy because preparatory to is
distributed across these two independent.

And as engineers you should probably
know that robustness in engineer IT

systems right is often
implemented through redundancy.

And feedback so actually an important

design principle underlying the Apollo
mission here but separate story.

Right so now each module persists in
preparatory activity and we are currently

exploring the mechanisms of how whole
preparatory activities maintain individual

molecules and this is ongoing work
let me show that to you briefly.

So.

Here's the problem right you've got
neurons in one side with different

selectivity if you inject information into
Iran briefly here different kinds of input

say during the instruction and part in our
behavior the information dissipates over

time skids of milliseconds really limited
by the membrane time constant but actually

quite quite a bit faster because of active
conduct so neurons so centrally it nicely.

Ation unconnected neurons are essentially
memory less in contrast preparatory

activity can last for
many seconds OK so now theorists have

been aware of this problem since the days
of Robinson in the eye movement.

At least thirty years and have come up
with many many models of how neurons might

maintain selective persistent take to
video of time skits or seconds for

the purposes of all kinds
of short memories and

they really fall into four or five
classes some little bit more exotic I've

just shown three kind of network models in
a cartoon fashion you could imagine that

neurons are arranged in a sequential
chain and hold memory in that

one set of neurons excite another set of
neurons so it's another set of neurons and

so on so forth Mark Goldman has
a nice model about that and

we think the dynamics of neurons is
not consistent in their response

to put to be Asians is not consistent with
this model and then there are two props

more plausible model it's there both
in the run of feedback dominated neural

networks that are dominated by excited for
networks excited for connections and

yet ball in both cases in the integrated
in the discrete attractive model of

principles are similar the time constant
in the neural networks are prolonged

by heat back and so there's dissipation
of through the membrane but

then you know you get
feedback from a neuron that

has similar queuing to give you that kick
back to keep you in that excited state and

integrators are distinguished from
attractors because they have a.

Continuous set of stable state and
they respond integrated manner

to graded input OK there is this create
attractors convert essentially continuous.

Putin sustained discrete outputs
here perhaps corresponding to

the movement directions it turns out and
one goal through all of

the experiments that really
are experiments point to this model

the discrete attractor model that's
underlying that in the mix of

let me show you two pieces of evidence
first the discrete attract a model really.

Predicts a funneling off with to
over time the delay Apoc right so

you've got some input sometimes
input is lower sometimes higher

noise in the system right and the if it T.

trajectories should
converge to discrete and

points also attract
will fix points during.

K.
and that's in fact what we see in the data

the set to fifty trajectories averaged in
a particular way I can't tell you about

the details but they seem to
converge to discrete a separate and

point during the delay in
this is been noted by others

OK perhaps and
there are several explanation for

this not all of the require to be discrete
attractor perhaps a more incisive and

stringent test again comes from these
perturbation experiments OK And

this experiment I like LOT OK I'll
go through this in some detail so

what you expect for these discrete
attractors is if you push the system

far but
not far enough not across the Senshi

the point that separates the two
Attractors and points of fixed points.

But also refer to this is
a separate trips in the language

of dynamic with systems some of
you may be taking a course in that

the tractor should recover to its
normal and point if you push.

The other hand the attractive cross the
network across a separate trips you should

recover to the other a tractor of
course with an error now in behavior

right and so this is what we did it so

this is and when I said we this
is really the ingenious work of.

A guy who did this experiment despite my

advice I didn't think this would work for
various reasons as usual I was wrong but

you know all of that stuff too so
anyway so here are the trajectories under

controlled conditions he devised
supportive ation that the collapse of

the selectivity is relatively little
effect on mean firing rate this is

sort of sucked away on some subtle
point and some try it's the perch of.

The correct trajectories recovered that
produces correct behavior on other

triads it's reject crease flip
OK right where trajectory

now locks in two snaps into the leftward
trajectory and vice versa with

an error trying exactly as predicted for
discrete attract the moderates and

I think this is the strongest experiment
that we have that really points

that discrete attracted to nomics
underlies prefatory activity.

And we have an idea of how
the basic circuit motif

looks like based on
analysis of connectivity.

And we think that neurons with
similar selectivity in dynamics

early that connect to each
of the preferentially and

recruit in addition in a rather to the FOS
biking into neurons in addition to the.

Promiscuous manner and it turns out
that moderates off this tie it.

The circuit motif actually produce
a tractor going to mix without

too much fine queuing of parameters.

I'm going to skip this is just a some
more details about the model and tell

you a little bit about more localization
I'm going to go through this very quickly

because I wanted to show you the last
bit which is just really more and.

So I'm going to return now very quickly
to the topic of globalization now

in the context of regional circuits so so
fire this big view of one brain region and

it turns out that's by far not
the whole story and it turns out that.

And we were interested in this because

again it's a question about
mechanisms right most modern of.

Preparatory activities short term
memory really rely on excited Tory

connections through
the horizontal connections so

localism K'naan cortical neurons
make a lot of axons locally and

people think well it's these connections
right within a lamp for example that

maintain memories and multiple lines of
evidence from the lab in the kit that it.

Plays a critical role in
maintaining persistent activity and

the experiment is very simple we looked at

other areas that are strongly connected
in a by interaction manner with.

Which one of these might be.

Critical for maintaining preparatory
activity and turns out that the small

strongly connected area of the cortical
areas have no effect every showing you

this from the other side of that right you
can blow out one side of preparatory to

the other side is maintained similar for
primary motor cortex and

some of the sensory cortex
in contrast if you silence.

That projects.

This absolutely stunning effect on
air mic to selectivity completely

cortex becomes coma told specifically and
the converse is also

true if you silence a limb and record
from fallenness it receives input from.

Selective it T.

and activity is lost in thought and so

we think that each side of this module
on each side that maintains preparatory

activity really relies on
feedback between and cortical.

OK that's required to mean it and
selectively having showing all

the pieces of data this was actually
published a couple months ago.

All right let me briefly very
very briefly summarize so

we've identified the key neural circuits
underlying motor preparation doing

motor preparation preparatory
activity conversions to and points.

And actions the shin noise refers to this
is essentially an initial condition for

movement in a space and
I think that's probably the right picture

reaching the appropriate initial condition
it's critical to produce fast and accurate

movements and I think these experiments
show this and preparatory activity is.

These initial these Fix Points
correspond to decease initial condition

does correspond to discrete Attractors and
Predator activities implemented in

modular and
redundant circuits All right so now so

far models of preparatory activity
are rhythmic in nature right so

they're basically a first sort of pop and
body minute hypothesis of how the dynamics

might be generated these kinds of then
it turns out that these kinds of network

more Teves that are moderate
relatively difficult to implement with

actually biologically plausible neurons
and I give you just one example.

This is a core tool that I made this
morning so it's a little rough but

the problem is one of in addition and
timescales are right so

if you have a neuron that spikes and then.

Evokes activity in inhibitory
population the feedback inhibition

comes in very quickly before the next
excited Tory Spike is fired and

that shuts off these networks and
so modelers have a workaround and

so what they do basically is they say
well look that's just makes an optic time

constants hundred milliseconds I don't
know if these synoptic time constants but

that bridge is right these the time
scale it's a feedback unit so

there are these crude has that are sort
of hidden in the in the methods so so

I think that the bio
physics really matters and

the cell types really matter and this
is something that of course we can take

seriously now because this is
really the era of so types and.

I think.

And so we in a collaboration with the on
brain in secure would have done for

the last few years and

dated all of the heavy lifting with
respect to seeing with sequencing done.

Essentially classification
of types in motor cortex and

what we find is I'm going to focus
only on two clusters this is

clustering of what we think there's
about forty neuron types of extremely

complicated we have a lot of work to
do this is a lot of that stuff but

on in the lab is informed by this
data this is not clustering by gene

expression right by transcript or makes
I'm going to just focus on two clusters.

Down here this corresponds
to the major output neurons

in layer five because I
think they're critical.

For maintaining predatory activity and

movement initiation irrelevant to the
story of maintaining persistent activity.

These two clusters correspond to
different projection classes and

this is done by combining
gene expression data with.

Structural analysis and this is a data
set from a shared project engineer.

Mike McConnell was one of the key
players of Post UK in the law.

And he's a single neuron
reconstructions for

neurons over late so
bodies are here and they fall into two

classes there are neurons here the yellow
and green project to the fallenness.

The magenta and blue project to
the medulla premotor centers

that also control oral facial movements
the that's that's really the premotor

centers of control the tongue are critical
for speaking and so on so forth so.

It's corresponding to these G
two gene expression classes and

they overlap in other places
like the superior curriculum.

Now here are here they are so
one project so

the thought one projects the middle up and
what is very exciting is that the.

Projection from the Father from
the cortex to the thoughtless projects

to the fallen is essentially ignores
coverage of neurons and the thought

is not excite the lawmaker to cure nucleus
neurons and this data from across.

Northwestern University is
a cartoon based on data.

So this could be a specialisation for
maintaining long time constants

in frontal cortico circuits and this is
what we're thinking about the problem now

there's a cortico Talum aquatic loop
that is excited Torrie only avoiding

this problem of rapid excited sort
of feedback that is a tendency.

Persistent activity.

There is a second specialization
which is probably why it when we

look through the gene expression data we
find that the thought of us projecting

neurons expressed at high levels and
exotic and India receptor

three A You know I used to work an M.B.A.
receptor that never heard of this thing

there are like five papers on it but it
has incredibly interesting properties for

the purposes of short term memory has
a long open Time hundred thirty seconds or

so and it lacks a magnesium BLOCK Does
everyone know what I'm uneasy block is so

in India receptors of voltage dependent
that's why they're not really that

useful for maintaining this persists and
they to a T.

and they're.

Because of magnesium block these
things lock them in easy and block so

they're opening addresses glutamate
binding and also they don't conduct so

they don't presume we don't use
an optic process of these so

they're perfect in terms of dynamics and

properties to maintain a long time
scale activity it's in fact it

is exactly the properties of the kind of
receptor is that moderates have you for

couple decades so those we don't know yet
if they're even at synapses

going to working model that these might
be critical to maintain preparatory.

But this is something in the age of
crisper that we can find out but

last piece of data.

So we hypothesize and these guys then
project two more centers would have

that this loop is somehow critical for
maintaining preparatory activity.

This kind of branch to initiate
movement movement period so

we did some so type specific recordings
in these two neuron types and

this appears to be the case so what we.

Doing is recording in a lemon and
exciting these different neuron types

by back firing that through
either from the middle or

the fallenness and then identifying
neurons that project here or there.

And when we now look at.

The persistence of selective and
thumb as projecting neurons we see

that they become selective remarkably
early maintain selective it

up to the Perry movement period through
their delay apocalypses so they so

there's memory short term memory and
thought of as projecting your own self as

hypothesized not in middle of projecting
your own Miller projecting neurons

the actual very rapid changes in selective
it after the go cue consistent with

triggering movements and these don't
show that projecting ones don't

show that kind of attitude so we think
that there is different output neurons and

there are five here to be specialized for
either maintaining motor plans and

initiate movements respectively and
so of course one of the key question

that we're interested in is how
this transformation from planning

to movement initiations achieve this
is one of the key question about how

information is gated from one
brain area to the other and

were very much working on that in general
I think in this circuit we now have

the opportunity to link neural circuits
to normal normal bio physics and to

neural computation and behavior and we've
sort of opened the black box right and

we know the road and brain region so
tide's how some of the dynamics

relate to behavior we can measure a key
state variables using imaging and.

Neurophysiology and up to genetics
allows us to shape state variables and

with methods like crisper would be able to
get at molecular mechanisms so we have.

Work cut out for ourselves over the next
few years and skip all of this and

then just thank all of
my collaborators and

let me highlight a few folks
in particular normally and.

Have started this work with me and
they're incredibly cool.

Guys there are no off cave
on he called my concern.

Carrying the torch on words he cause
on the job market he's absolutely

brilliant OK this guy is also learning
on the job market the sculls.

Serato will take a while and

then very valuable collaborators in
matters of theory and data analysis.

Druckman son Ramani the runs
of fun to learn and

other collaborators whose work I
only was able to skim and thanks for

your attention and
I'll take questions if there are no.

Thank.

You.

Yeah yeah Yeah absolutely so
so this is critical right so

in all of these tasks the dynamics of
the activity to some extent reflects

the dynamics and the and the and
the dimensionality of the behaviors so

the dimensionality is low because
behavior is low so and we think that.

Tractors are learned and
that if we train additional endpoints

into this we'll get additional
tractors and this is the kind of

thing that we're trying to that we're
doing our side of the damage now T.

we're not in general general be.

As low dimensional so
sorry I got Mouli has done some analysis

in on the various data set including.

Noise status at this is a reaching
task and he's come to that conclusion.

You know we have an intuition that
that might be the case but so

it has put more quantitative meat on this
notion that you didn't mention ality in

these in these modern dynamics
through to reflect spent behavior.

Yeah so that's a good that's so so
whether or not you get discrete attractors

that's a good question so
this is something that is.

So what we're what we're doing is
training in additional targets and

trying to see when are that
that looks this is ongoing

corresponds to additional discrete
attractors I think really the question is

what what what about short term
memories that now require.

You know require representational
continues variable and

essentially flexible.

Use of that information like
in a working memory task and

that's also something we're
interested in I would imagine.

Tractor dynamics would
not deal with that in a.

Parsimonious way so yeah that would be
an interesting point for comparison and

you can you know perhaps use the same
kind of unleashed the same kind of

experimental arsenal to look
at the underlying circuit.

Implementation So
we haven't done that yet young.

One.

Yeah.

I don't know about the decision and

stuff you know I don't really know I think
I think I personally I like to talk about.

Motor preparation and
I think decision the variable are probably

very closely linked to motor preparation
and I don't know if a single experiment

that is read was to the east to do
this sort of a bigger picture I would

get in huge lot of trouble in certain
departments with that kind of statement.

But so on spread the word.

But you're bigger you're really a bigger
question is a mechanistic want right so so

we we you thinking about
our NAND specifically And

so these are recurrent neural networks and
so

we think you can train
our ends to look like

discrete attractors effectively or
like integrators and

in fact if you if you just impose this
they tend to want to be integrators.

And so it's kind of.

And then others so so I think and

then I think I think the principle
turned out to be very similar.

Although you don't have to go in and

reverse engineer these things and
you know.

It can be complex and in and off itself
of but so depending on what you train.

So if you if you train them
just to produce to fix points

they're produced discrete
attractors the So So

I I think those are kind of Arar's
us in some sense and the way.

The details and how you treat
at least in our experience and

we're not as good at this is we don't
understand these things as deeply

as some others that you kind
of you get what you train and

so I I like this more this kind of
I'll agree with make approach and

then you know you just sort of go and
sort of.

Ask you know what is this
dynamic not consistent with and

that's that's you know
if you go go from there.

So that that's that's something
we don't entirely know so

there's a question of whether or not
this area is is specific to oral facial

movements in the context of motor planning
and we we don't know the answer to that so

there's just very little work that has
gone into this kind of detail to really

look for the relevant brain regions
in unbiased manner and you know

it's something that someone should really
do in a reaching task for example or so.

You know early we done some very
preliminary stuff with kind of directional

navigation were it looked like it did
play a role but it's too early to tell.

You I unfortunately don't know the answer.

But I think we have to vacate
the room is a writer who thinks.