nature neuroscience • volume 1 no 4 • august 1998 261

news and views

Anyone who was led tobelieve that catching abird simply requiressprinkling salt on its tailcan probably testify thatthis method does notwork well. Whether ornot salt on the tail has aparalyzing effect, livebirds rarely give you theopportunity to test themethod. In fact, mostanimals have evolvedmany parallel warningsystems to escape suchundesirable encounterswith predators. The smellof coyote urine makeswild mice freeze, thesound of bat calls makesflying crickets dive, andthe sight of an approach-ing car makes humansrun (or brake). Amongwarning signals, thosecoming from movingpredators present a com-plex challenge to thebrain: how can theirdynamic physical charac-teristics indicate with little ambiguitythat they represent a looming danger?Work by Sun and Frost1 on page 296 ofthis issue of Nature Neuroscience pro-poses a set of solutions expressed by thedynamic responses of specialized visualneurons in pigeons. What makes thiswork all the more interesting to us isthat it indicates, first, that several dif-ferent computations—each with its ownadvantages and disadvantages—are car-ried out in parallel in the same region

na. From these variables,an observer can derivesome of the characteris-tics of the object’smotion, as well as predictthe time of impendingcollision. The big ques-tion is, how does thebrain behind the exposedretina do it?

One possibility is thatneurons (or neuronal cir-cuits) compute ρ(t), therate of angular expan-sion, thus indirectlytracking object approach.This solution is reason-ably simple and couldlead to an appropriateavoidance commandwhen the firing rate ofthese neurons crosses acertain threshold. It alsohas an advantage:because large objectsstart to appear big earli-er than small ones, thethreshold will be crossedearlier during approachof large objects, leaving

more time for escape. The solution hasits downside too, in that the rate ofexpansion increases faster as collisionapproaches. If this parameter (the rateof expansion) is represented by a firingrate, the neuron’s firing may saturateclose to the critical time.

To avoid this, Mother Nature founda solution: the fast expansion can beslowed down by dividing the rate ofangular expansion by an exponentialfunction of the object’s angular size[η(t) = θ’(t) / eα.θ(t)] (Fig. 1c). Whenthe object is far away, the growth of thenumerator dominates and η(t) (or thefiring rate of the neuron representingit) increases. When the object comesnearer, the denominator gains relativeinfluence because of its exponentiation

of the brain and, second, that one ofthese solutions, despite its apparentcomplexity, is found unchanged in verydifferent animals (birds1 and insects2).

An approaching object on a collisioncourse projects an expanding image onthe retina (Fig. 1a). If the approachvelocity is constant, the angle θ(t) sub-tended by the object grows nonlinearlyin a near-to-exponential fashion. Thiscan be seen in Fig. 1a, where the growthof θ is greater over the late than over theearly half of the object’s approach. Sim-ilarly, the rate at which this angleexpands (ρ(t) = θ’(t)) itself increasesnonlinearly (Fig. 1b). For a given objectapproach velocity, specific attributes ofangular expansion (size, velocity, accel-eration) are thus projected on the reti-

Collision-avoidance: nature’s manysolutionsGilles Laurent and Fabrizio Gabbiani

How does the brain sense looming danger? A new study shows that specialized visualneurons in pigeons carry out several different computations in parallel to analyze signalsfrom approaching objects such as predators.

Gilles Laurent and Fabrizio Gabbiani are at theDivision of Biology, California Institute ofTechnology, 139-74, Pasadena, California 91125, USAe-mail: GL (laurentg@cco.caltech.edu) orFG (gabbiani@klab.caltech.edu)

θ

v

d

-3 -2 -1 00

10

20

30

40

50

Time to collision (sec)

An

gul

ar s

ize

(deg

)

0

100

200

300

400

500

An

gul

ar s

pee

d (

deg

/sec

)

-3 -2 -1 00

0.5

1

1.5

2

2.5

3

Time to collision (sec)

1/Ti

me

to c

olli

sio

n (

sec-1

)

0

0.2

0.4

0.6

0.8

1

h(t

) (a

rbit

rary

un

its)η(t)

τ(t)

1/τ(t)

θ(t)

ρ(t)

a

b c

Fig. 1. Elementary kinematics of object approach on a collision course (a) Schematics of a typical looming experiment1. An object of fixed size dapproaches the eye at a constant velocity v. During approach, the angle θ(t)subtended by the object and its rate of increase [ρ(t) = θ’(t)] both grow non-linearly. (b) Time course of θ(t) and ρ(t) during approach (v = 300 cm per s,d = 30 cm). (c) Corresponding time course of the functions τ(t), τ–1(t) andη(t) (with α = 16, see eq. 2 in ref. 1).

Time to collision (s)

a

b

An

gu

lar

size

(d

eg)

Time to collision (s)

An

gu

lar

spee

d (

deg

/s)

1/ti

me

to c

olli

sio

n (

s-1)

c

η(t)

(ar

bit

rary

un

its)

© 1998 Nature America Inc. • http://neurosci.nature.com©

199

8 N

atu

re A

mer

ica

Inc.

• h

ttp

://n

euro

sci.n

atu

re.c

om

262 nature neuroscience • volume 1 no 4 • august 1998

news and views

and slows down the increase in firingrate. Another advantage is that, at a crit-ical time preceding collision, η(t) reach-es a peak. This peak occurs when (andthus signals that) the object subtends aparticular angle, given by 2tg-1(2/α) ,constant for a given neuron. There aretwo problems associated with η , how-ever. First, to know that this angle hasbeen reached, downstream circuits needto detect a peak firing rate. This is nottrivial, and because peak detectionrequires comparing successive values,this operation requires time, when timeis in short supply. Second, the peak sig-nals an angle, not an actual size; it thusconfounds a large object that is fartheraway with a small object that is near.Less time would therefore be availablefor escape from a small and rapidlyapproaching object than from a large,slow moving one.

Here also, there is a potential solu-tion. It consists of calculating yet anoth-er variable, the time-to-contact τ(t) orits inverse (Fig. 1c). Tau is relatively easyto compute when θ(t) is small and whenthe approach velocity is constant:τ ≈ θ(t)/θ’(t). This measure, introducedby Gibson3 and studied behaviorally indiving birds by Lee and Reddish4, hasthe advantage that it is independent ofthe object size or approach velocity. Tauor 1/τ(t) gives a running value of the

time before collision. By setting anappropriate threshold, it becomes pos-sible to trigger a motor reaction at aconstant delay prior to the anticipatedcollision. The downsides of the τ com-putation are that it provides no infor-mation about object size or velocity andthat the mathematical approximationτ ≈ θ/θ’ is not valid when θ is large.

Remarkably, each of these three pot-netial solutions is reflected in the prop-erties of neurons in nucleus rotundus ofpigeons1, an area homologous to themammalian inferior caudal pulvinar5—a thalamic nucleus with visual inputsfrom the superior colliculus and whichprojects to occipital, parietal and tem-poral cortices. Sun and Frost1 identifiedthree groups of neurons that respond toapproaching objects on a collisioncourse by comparing the dynamics oftheir firing rates with kinematic func-tions such as θ(t), ρ(t) or η(t) for dif-ferent object sizes and approachvelocities. (See Fig. 2 for a summary oftheir methods.) One group of neuronsshows firing profiles that are bestdescribed by a ρ computation. A secondgroup shows peaked firing profiles, bestfitted by a η computation. This is par-ticularly interesting to us because the ηalgorithm was first derived2 to describethe responses of DCMD, a looming-sen-sitive neuron in locusts6–8. To find such

a remarkably similar solution in suchdistant species (which interestingly havesimilar predators) supports the idea thatsimilar problems engender similar com-putational solutions. Finally, Sun andFrost describe a third group of neuronswhose onset of activity during approachis independent of object size or veloci-ty, suggesting a τ-style computation.The existence of this last class of neu-rons was previously reported by Frost’sgroup9. The properties of these neuronswere confirmed in this report and ana-lyzed further, allowing their clear dis-tinction from the other two neuronaland computational clusters.

Why are these results important?First, they focus on the dynamics of neu-ronal responses—rather than meanresponses—as relevant neuronal signals.An early attempt at this was made byRind and Simmons8 in their study oflocust DCMD responses. Second, theresults indicate that the brain recon-structs object approach using several par-allel (and possibly serial) computations.Each one provides a different piece ofinformation about the state of the envi-ronment, and the animal thus presum-ably makes an informed decision on thebasis of these different inputs. But how?It will be interesting to study how down-stream circuits interpret these parallelmessages, so as to make the best motor

Fig. 2. To classify neurons into different groups, Sun and Frost1 com-pared the instantaneous firing rate (PSTHs) of their recorded neu-rons with different functions of time such as θ(t), ρ(t), and η(t) (in a1,b1 and c1, respectively) for different values of d/v (d1/v1 = 0.02 s,d2/v2 = 0.1 s, d3/v3 = 0.2 s in a-c). Each function also predicts specificrelations between the stimulus and the PSTH. (a) If the relationshipfollows θ(t), the time at which a given angular threshold is crossed(a1) should be linearly related to d/v (a2). (b) If the relationship fol-lows ρ(t), the time at which a given angular velocity threshold iscrossed (b1) should be linearly related to (d/v)1/2 (b2). (c) If the rela-tionship follows η(t), the time of the peak in the PSTH (c1) shouldbe linearly related to d/v (c2). Sun and Frost’ s results1 indicate thepresence of two classes corresponding to (b) and (c) but none to (a).In addition, a third class has a constant firing threshold, consistentwith a τ computation.

a1 a2

d1/v1

d2/v2

d3/v3

-3 -2 -1 00

5

10

15

20

An

gul

ar s

ize

(deg

)

0 0.05 0.1 0.15 0.20

0.5

1

1.5

2

2.5

d/v (sec)

T th

res

(sec

)

θ(t)

Time to collision (sec)

b1 b2

-3 -2 -1 00

5

10

15

20

25

An

gul

ar s

pee

d (

deg

/sec

)

0.1 0.2 0.3 0.4 0.50

0.5

1

1.5

2

1/2 1/2

T th

res

(sec

)

ρ(t)

c1 c2

-3 -2 -1 00

0.2

0.4

0.6

0.8

1

η(t)

(ar

bit

rary

un

its)

0 0.05 0.1 0.15 0.20

0.5

1

1.5

2

T pea

k (s

ec)

η(t)

d/v (sec)Time to collision (sec)

( ) ( )

An

gu

lar

size

(d

eg)

a1

Time to collision (s)

Tth

res

(s)

d/v (s)

a2 b2

c1 c2

An

gu

lar

spee

d (

deg

/s)

b1

Tth

res

(s)

Time to collision (s) (d/v)1/2 (s1/2)

Time to collision (s)η(

t) (

arb

itra

ry u

nit

s)

Tp

eak

(s)

d/v (s)

© 1998 Nature America Inc. • http://neurosci.nature.com©

199

8 N

atu

re A

mer

ica

Inc.

• h

ttp

://n

euro

sci.n

atu

re.c

om

news and views

nature neuroscience • volume 1 no 4 • august 1998 263

decision (for example, “duck, but not tooearly, so as to prevent course correctionby predator”). Does one signal dominatethe others (see ref. 9)? If so, under whatcircumstances? Or does some combina-tion of these signals guide behavioralresponses? It will also be fascinating tostudy the cellular mechanisms by whichthese different computations are carriedout. Do the ρ neurons for example, pro-vide inputs to the τ and η neurons,allowing size and velocity signals to becombined? If so, how are these opera-

3. Gibson, J. J. The Ecological Approach to VisualPerception (Houghton Mifflin, Boston, 1979).

4. Lee, D. N. & Reddish, P. E. Nature 293,293–294 (1980).

5. Karten H. J., Cox, K. & Mpodozis, J. J. Comp.Neurol. 387, 449–465 (1997).

6. Palka, J. J. Insect. Physiol. 13, 235–248 (1967).

7. Schlotterer, G. R. Can. J. Zool. 55, 1372–1376(1977).

8. Rind, F. C. & Simmons, P. J. J. Neurophysiol.68, 1654–1666 (1992).

9. Wang, Y. & Frost, B. J. Nature 356, 236–238(1992).

tions implemented biophysically? If theτ neurons really fire when the variable τcrosses a given threshold, how is thisthreshold set and held constant? Theseare some of the many interesting ques-tions that remain to be answered, butBarrie Frost and colleagues are getting uscloser to this target.

1. Sun, H. & Frost, B. J. Nature Neuroscience 1,296–303 (1998).

2. Hatsopoulos, N., Gabbiani, F. & Laurent, G.Science 270, 1000–1003 (1995).

Neurogenic control of thecerebral microcirculation: isdopamine minding the store?Costantino Iadecola

Cerebral blood flow is highly regulated by neural activity.New anatomical and functional evidence suggests thatdopamine neurons may play a key role in this process.

It has long been known that cerebralblood vessels receive abundant projec-tions from central and peripheral neu-rons, but the functional significance ofthis innervation has remained elusive.A new study by Krimer and colleagueson pages 286–289 of this issue of NatureNeuroscience provides evidence that cen-tral dopaminergic neurons are uniquelypositioned to control the cerebralmicrocirculation and that they may par-ticipate in the regulation of cerebralblood flow by neural activity.

The ability to map changes in cere-bral blood flow produced by neuralactivity using functional imaging hasproven to be one of the most powerfultools for localization of function in thebehaving human brain1. Yet the funda-mental mechanisms linking neuralactivity to cerebral blood flow are notfully understood. The brain, perhapsmore than any other organ of the body,has the intrinsic ability to regulate itsown blood flow with a high degree of

spatial and temporal precision2. Theamount of flow that each brain regionreceives is directly related to the func-tional activity of that region. Thus, if theneural activity increases, for example inspecific areas of the cerebral cortex dur-ing somatosensory or visual stimuli,cerebral blood flow to the activatedregions also increases. Cerebral arteriestravel on the surface of the brain (pialarteries) and then enter the brainparenchyma (penetrating arterioles),giving rise to smaller arterioles and cap-illaries, which supply the tissue withnutrients and remove waste. Cerebralblood vessels have the ability to contractand relax in response to a wide varietyof stimuli, thereby decreasing orincreasing flow to the different areas ofthe brain.

The mechanisms of this activity-dependent regulation of blood flow havebeen investigated for many decades. Awidely accepted hypothesis, articulatedby Roy and Sherrington in 1890, is thatworking neurons release vasoactiveagents in the extracellular space, whichreach blood vessels by diffusion and pro-duce relaxation of vascular smooth mus-cles3. However, diffusion of vasoactivemetabolites, such as nitric oxide, K+ andH+ ions and adenosine from active neu-

rons cannot account in full for the rapid-ity and spatial definition of the increas-es in cerebral blood flow produced byneural activity. It was therefore proposedthat selected neurons project to cerebralblood vessels and regulate cerebral bloodflow by influencing vascular diameterdirectly. Although it is well known thatcerebral blood vessels are closely associ-ated with neural processes, there hasbeen a long-standing controversy as towhether this proposal is correct becauseclear morphological and functional evi-dence linking central neural processes tocontractile elements of the cerebral ves-sel wall was lacking2.

Now Krimer and colleagues providenew evidence supporting a direct role ofcentral vascular terminals in the dynam-ic regulation of the cerebral microcircu-lation. Using immunocytochemicaltechniques to visualize dopaminergic ter-minals both at the light and electronmicroscopic level, they demonstrate thatprocesses from central dopaminergic neu-rons terminate in close contact with pen-etrating arterioles and cerebral capillariesin the cerebral cortex. The density of thedopaminergic vascular innervation variesregionally, being greatest in dopamine-rich regions of the frontal, sensorimotorand entorhinal cortices. The dopaminer-gic processes terminate either directly onthe vascular basal lamina or on perivas-cular astrocytic end-feet. In arterioles, theterminals are located on penetrating arte-rioles, which are important in the distri-bution of cerebral blood flow. At thecapillary level, the terminals are almostinvariably located in close proximity topericytes, which are contractile cellsembedded in the microvascular basal lam-ina. Krimer and colleagues also show thatmicroapplication of dopamine in thevicinity of cerebral microvessels by ion-tophoresis, produces vasoconstriction inapproximately 50% of the microvessels

Costantino Iadecola is at the Laboratory ofCerebrovascular Biology and Stroke,Department of Neurology, University ofMinnesota, Box 295 UMHC, 420 DelawareStreet S.E., Minneapolis, Minnesota 55455, USAemail: iadec001@tc.umn.edu

© 1998 Nature America Inc. • http://neurosci.nature.com©

199

8 N

atu

re A

mer

ica

Inc.

• h

ttp

://n

euro

sci.n

atu

re.c

om