DECODING THE FUNCTIONAL NETWORKS OF CEREBRAL CORTEX

SIMONS COMPUTATIONAL THEORIES OF THE BRAIN APRIL 18, 2018

226

J. Physiol. (I959) I47, 226-238

SINGLE UNIT ACTIVITY IN STRIATE CORTEX OFUNRESTRAINED CATS

BY D. H. HUBEL*From the Department of Neurophysiology, Walter Reed Army Institute

of Research, Walter Reed Army Medical Center,Washington 12, D.C., U.S.A.

(Received 15 December 1958)

A beginning has recently been made in recording single neurone activity fromanimals with chronically implanted electrodes (Hubel, 1957a; Gusel'nikov,1957; Ricci, Doane & Jasper, 1957; Strumwasser, 1958). These methodseliminate anaesthetics, paralysing drugs, brain-stem lesions, and otheracute experimental procedures. They make it possible to record electricalevents in the higher central nervous system with the animal in a normal state,and to correlate these electrical events with such variables as waking state,attention, learning, and motor activity.The present paper describes a method for unit recording from the cortex of

unanaesthetized, unrestrained cats, and presents some observations from thestriate cortex. The objectives have been (1) to observe maintained unit acti-vity under various conditions such as sleep and wakefulness, and (2) to findfor each unit the natural stimuli which most effectively influence firing. Of400 units observed, some 200 are presented here because of their commoncharacteristics. Since there is reason to believe that the remainiing 200 unitswere afferent fibres from the lateral geniculate nucleus, these will be de-scribed in a separate paper. A preliminary account of some of this work hasbeen given elsewhere (Hubel, 1958).

METHODSUnit recordings in unrestrained animals were made with micro-electrodes held in a positionerwhich was anchored rigidly to the skull during recordings, and removed between recordings.A rigidly implanted peg adapted from the design of Ricci et al. (1957) held the micropositioner atthe time of recording (Text-fig. 1). The peg was made of the plastic Kel-F (fluorocarbon polymermade by Minnesota Mining and Manufacturing Company, St Paul, Minn.). It was hollow and was

* Present addreas: Neurophysiological Laboratory, Department of Pharmacology, HarvardMedical School, Boston 15, Massachusetts, U.S.A.

STRIATE UNIT RECORDINGS 229(1941) this should be close to the region where central vision is represented. The brains of all catswere examined grossly and histologically to confirm the area and to assess damage. In mostbrains grey matter appeared normal to Nissl stain, except for some round-cell infiltration whichwas confined to the leptomeninges, electrode tracks, and perivascular spaces.The depths at which units were recorded were not accurately known because of the sharp

curvature of the lateral gyrus, dimpling of the surface of the brain at the time of penetration, anddifficulty, due to overlying cerebrospinal fluid, in knowing when the electrode made contact withthe surface of the cortex. However, one could be sure that most units were less than 2 mm deep-the approximate thickness of the cortical grey matter-since the total distance travelled by theelectrode was usually no more than this.

In six cats electrolytic lesions were made by passing 5 pA for 5 sec through the electode tip(electrode negative). Brains were perfused with saline followed by 10% formalin within 12 hr ofmaking lesions. Serial paraffin sections 15 i thick were stained with cresvviolet. Lesions were ofremarkably uniform size and configuration (P1. 2), and consisted of a dark-staining core about351A in diameter surrounded by a pale halo. In larger lesions there was a clear area in the centre.The size of the halo was 100-200,u in diameter, and was a great help in finding the lesion in alarge number of serial sections.When lesions were made in several closely spaced penetrations it became necessary to disting-

uish one penetration from the other. This was done by making a different number of lesions alongeach track, using the number of lesions to identify the penetration. Microscopic examination ofelectrolytic lesions such as that shown in P1. 2 further supported the conclusion that most recordswere from the grey matter of lateral gyrus.

Since these electrodes can record from single myelinated fibres (Hubel, 1957b), it is not certainthat all records were obtained from cell bodies; some may have been from axons of cortical cells orfrom afferent fibres. Electrophysiologically it is not easy to distinguish cell-body from axonrecords. Although some uncertainty remains in this respect, it is believed that most records werefrom cell bodies.

RESULTS

Background activityIn the unrestrained preparation most units showed activity in the absence ofintentional stimulation on the part of the observer. As the cat looked about,spurts and pauses in firing were seen to accompany eye movements. When theeyes were closed either passively by the observer or by the cat, firing usuallypersisted, although it was generally less active. Even when the room wasmade completely dark most units continued to fire.

Cortical unit discharges were very irregular. In the more active units firingoften occurred in bursts separated by silent intervals (Text-figs. 2-4). Occa-sionally bursts recurred rhythmically for short periods, as is seen in Text-fig. 2. Even in units in which grouping was not obvious, firing was not nearlyas regular as in other systems such as spinal motoneurones and peripheralstretch receptors.

Satisfactory studies comparing activity in sleeping and waking states weremade in twelve units. One typical change during arousal is illustrated inText-fig. 3. Bursts were either smoothed out, giving a marked increase inregularity of firing, or they were reduced in length while occurring more often.In either case there was no obvious change in the over-all rate of firing.

VISUAL CORTEX DISPLAYS ACTIVITY NOT DIRECTLY TIED TO VISUAL STIMULI

▸ Extend theories of cortical computation from the average case to single trials using network based analytical tools

▸ Incorporate a more comprehensive sampling of the network to include unbiased sample of neurons

GOALS

OUTLINE

▸Mouse V1 & high speed two-photon imaging

▸What are functional networks?

▸Functional networks accurately predict neuronal activity

▸Higher order structure in functional networks

▸Assemblies

Window over V1

Top down view

Head bar

OUR EXPERIMENTAL SET-UP▸ Awake behaving (ambulating) animals

RETINOTOPY TO CONFIRM VISUAL CORTEX

Window over V1

Top down view

Head bar

Intrinsic Imaging Two Photon Imaging

IN VIVO 2P SOMATIC IMAGING OF VISUALLY EVOKED ACTIVITY

CORTICAL V1 MICROCIRCUIT DYNAMICS IMAGED WITH HOPS

MULTINEURONAL RESPONSES TO DRIFTING GRATINGS

R

grating direction

(s)

normalized activity (A

.U.)

50

100

150

neur

ons

50 100 150 200 250

10 cm/s Running speed

REPRESENTATIVE TUNED AND UNTUNED RESPONSES

grating direction

Cell 1

Cell 2

Cell 3

Cell 4

Cell 5

TUNED NEURONS SHOW VARIABLE SINGLE TRIAL ACTIVITYtri

als

time

90

270

180 0

norm. fluorescence

trial

s

time

90

270

180 0

norm. fluorescence

trial

s

1 s

1

30

TUNED NEURONS SHOW VARIABLE SINGLE TRIAL ACTIVITY

TUNING DOESN’T PREDICT SINGLE TRIAL RESPONSES VERY WELL

0 0.2 0.4 0.6 0.8 1variance explained

0

5

10

15

20

prob

abilit

y de

nsity

tuned (2613 / 4535)untuned (1922 / 4535)

OUTLINE

▸Mouse V1 & high speed two-photon imaging

▸What are functional networks?

▸Functional networks accurately predict neuronal activity

▸Higher order structure in functional networks

Network:Amathema)cal representa)onofa real-worldcomplex system defined by a collec)on of nodes(ver)ces)andlinks(edges)betweenpairsofnodes.

BUILDING A FUNCTIONAL NETWORK USING PAIRWISE PARTIAL CORRELATION

Directionality = correlogram lag

zero lagi j

tmax

within movie activityi,j

remaining movie averagei,j

within movie averagek≠i,j

50%20 sec

Edge weight = partial correlationw = 0.51

FUNCTIONAL NETWORKS CONTAIN EDGES BETWEEN TUNED AND UNTUNED NEURONS

neur

on idtuned untuned

edge

wei

ght

0

0.2

0.4tu

ned

untu

ned

neuron id

FUNCTIONAL NETWORKS REFLECT TUNING IN THE POPULATION

neur

on id

tuned untuned

edge

wei

ght

0

0.2

0.4

tune

dun

tune

d

neuron id

0.04

0.06

0.08

0.1

edge

wei

ght

tuned - tunedtuned - untuneduntuned - untuned

difference in preferred direction (o)0 90 180

OUTLINE

▸Mouse V1 & high speed two-photon imaging

▸What are functional networks?

▸Functional networks accurately predict neuronal activity

▸Higher order structure in functional networks

MODELING NEURON RESPONSES USING FUNCTIONAL NETWORKS

neur

on id

tuned untuned

edge

wei

ght

0

0.2

0.4

tune

dun

tune

d

neuron id

rescale

weightsinput neuron activity

∑

W1

W2

Wn

50% ΔF 20 sec

rescale

weightsinput neuron activity

∑

W1

W2

Wn

50% ΔF 20 sec

predicted activity / measured activity

20 sec

50% ΔF

ACCURATE PREDICTION OF MOMENT TO MOMENT ACTIVITY USING FUNCTIONAL NETWORKS

FUNCTIONAL NETWORKS PROVIDE NEAR OPTIMAL PREDICTIONS OF SINGLE TRIAL RESPONSES

predicted activity / measured activity

50%20 sec

optim

al w

eigh

ts (M

SE

)

graph weights (MSE)10-2

10 -3

10 -2

10 -1

10-1

FUNCTIONAL NETWORKS ALSO PREDICT TUNING

00

1predicted tuning vector

(cosine similarity)

200

400

600ne

uron

cou

nt

0.2 0.4 0.6 0.8

0

30

6090

120

150

180

210

240270

300

330

Cosine similarity = 1

0

30

6090

120

150

180

210

240270

300

330

Cosine similarity < 1

50 100 150 200 250 300 350neurons

0.4

0.5

0.6

0.7

0.8

0.9

popu

latio

n va

rianc

e ex

plai

ned

POPULATION SIZE UNDERLIES PREDICTION ACCURACY

LARGE WEIGHTS CONTRIBUTE DISPROPORTIONATELY TO PREDICTION ACCURACY

% to

tal M

SE

0.200 0.4 0.6 0.8fraction weight removed

0.4

0.5

0.6

0.7

0.8

0.9

1

% to

tal M

SE

0.2 0.4 0.6 0.8fraction edges removed

0.4

0.5

0.6

0.7

0.8

0.9

1

weakest first

strongest firstrandom

weakest first

strongest firstrandom

0 0.2 0.4 0.6 0.8edge weight threshold (quantile)

1

10re

cipr

ocal

pro

babi

lity

(fold

ove

r ran

dom

)within tunedwithin untunedbetween tuned & untuned

RECURRENT CONNECTIONS ARE BIASED TOWARD LARGE EDGE WEIGHTS

OUTLINE

▸Mouse V1 & high speed two-photon imaging

▸What are functional networks?

▸Functional networks accurately predict neuronal activity

▸moment to moment prediction

▸Higher order structure in functional networks

BEYOND PAIRWISE

Directed

middle-man

fan-out

undirected

fan-in cycle

TRIPLET MOTIF STRUCTURE IN FUNCTIONAL NETWORKS

middle-man

fan-out

undirected

fan-in

cycle

0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8clustering coefficient(fold over Random)

TRIPLET MOTIF STRUCTURE UNDERLIES PREDICTION ACCURACY

0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8clustering coefficient

(fold over ER)

cycle

middle

fan-in

fan-out

undirected

0.2

0.4

0.6

0.8

1

varia

nce

expl

aine

d

-0.5 0 0.5 1(z-score)

middle

Dominance of Motif

CONCLUSIONS

▸Neurons are variable making prediction of single trial activity from tuning properties difficult

▸Functional networks provide near optimal predictions of activity in individual neurons

▸And predict tuning

▸Triplet correlations are predictive of more than 90% of a moment to moment activity

▸ Data demonstrates a loose coalition of neurons that co-vary with one another and consequently are predictive of one another

▸ Multineuronal activity shows pairwise timing differences as indicated by the fact that the majority of entries in the matrix are asymmetric

▸ But many unknowns

ASSEMBLIES

ACKNOWLEDGMENTS

Lab Alumni

Brendan Chambers Alex Sadovsky Peter Kruskal Melissa Runfeldt Lucy Li SJ Weinberg Veronika Hanko Lane McIntosh Suchin Gururangan Charles Frye Alexa Carlson Caroline Heimerl Isabella Penido Areknaz Khaligian Audrey Sederberg

Current Lab

Joe Dechery Vaughn Spurrier Maayan Levy Peter Malonis Zania Zayyad Kyle Bojanek Subhodh Kotekal Isabel Garon Carolina Yu Harrison Grier Friederice Pirschen Anne Havlik

FACCTS