Top-Down Feedback Controls Spatial Summation and Response Gain in Primate Visual Cortex

In the cerebral cortex, sensory information travels along feedforward connections through a hierarchy of areas, which, in turn, send a denser network of feedback connections to lower-order areas. Feedback has been implicated in attention, expectation, and sensory context, but the cellular mechanisms underlying these diverse feedback functions are unknown. Using specific optogenetic inactivation of feedback connections in the primate visual cortex, we have identified the cellular mechanisms of feedback-mediated modulations of early sensory processing. Specifically, we found that feedback modulates receptive field size, surround suppression and response gain, similar to the modulatory effects of visual spatial attention. A recurrent network model captured these effects. These feedback-mediated modulations allow higher-order cortical areas to dynamically regulate spatial resolution, sensitivity to image features, and efficiency of coding natural images in lower-order cortical areas.


24
In the cerebral cortex, sensory information travels along feedforward connections through a 25 hierarchy of areas, which, in turn, send a denser network of feedback connections to lower-order 26 areas. Feedback has been implicated in attention, expectation, and sensory context, but the 27 cellular mechanisms underlying these diverse feedback functions are unknown. Using specific 28 optogenetic inactivation of feedback connections in the primate visual cortex, we have identified 29 the cellular mechanisms of feedback-mediated modulations of early sensory processing. 30 Specifically, we found that feedback modulates receptive field size, surround suppression and 31 response gain, similar to the modulatory effects of visual spatial attention. A recurrent network 32 model captured these effects. These feedback-mediated modulations allow higher-order cortical 33 areas to dynamically regulate spatial resolution, sensitivity to image features, and efficiency of 34 coding natural images in lower-order cortical areas. 35 3 In addition to well studied bottom-up feedforward inputs, the visual cortex receives a much 36 denser network of feedback inputs from higher-order cortical areas 1 whose role remains 37 hypothetical. Feedback has been implicated in several forms of top-down influences, such as 38 attention 2,3 , expectation 4 and sensory context 5,6 , which affect sensory processing in diverse ways. 39 For example, visual spatial attention, one of the most studied instances of top-down influences, 40 has been shown to modulate neuronal response gain 2,7 , surround suppression 8 and receptive field 41 (RF) size 9 . In this study we have asked whether feedback connections can mediate such diverse 42 effects. 43 To determine the cellular mechanisms underlying the influence of cortical feedback on 44 sensory processing, we asked whether inactivating feedback from the secondary visual area (V2) 45 alters RF size, surround suppression and response gain in the primary visual cortex (V1). 46 Surround suppression is the property of V1 neurons to reduce their response to stimuli inside 47 their RF when presented with large stimuli extending into the RF surround 10-18 . This is a 48 fundamental computation throughout the visual cortex, thought to increase the neurons coding 49 efficiency [19][20][21][22] , to contribute to segmentation of objects boundaries 21 , and to be generated by 50 feedback connections 5,6 . However, the role of feedback in surround suppression and response 51 gain remains controversial. Inactivation of higher-order cortices using pharmacology, cooling or 52 optogenetics has produced weak reduction in surround suppression in some studies [23][24][25] , but only 53 reduction in response gain in other studies [26][27][28][29] . One problem with these previous studies is that 54 these inactivation methods suppressed activity in an entire cortical area; thus, the observed 55 effects could have resulted from indirect pathways through the thalamus or other cortical areas. 56 Moreover, these approaches did not allow fine control of inactivation levels, thus precluding 57 potentially more physiologically relevant manipulations. To overcome the technical limitations 58 of previous studies, we have used selective optogenetic inactivation of V2-to-V1 feedback 59 terminals, rather than direct inactivation of the entire V2, while measuring spatial summation and 60 surround suppression in V1 neurons using linear electrode arrays. 61 62 63 RESULTS 64 65 Specific Optogenetic Inactivation of Feedback Connections 66 To express the outward proton pump Archaerhodopsin-T (ArchT) 30 in the axon terminals of V2 67 feedback neurons, we injected into V2 of marmoset monkeys a mixture of Cre-expressing and 68 Cre-dependent adeno-associated virus (AAV9) carrying the genes for ArchT and green 69 fluorescent protein (Fig. 1a,c; see Methods). This viral vector combination was used because in 70 pilot studies we found it produces selective anterograde infection of neurons at the injected V2 71 site, and virtually no retrograde infection of neurons in V1 (Fig. 1d). Intrinsic signal optical 72 imaging was used to identify the V1/V2 border ( Fig. 1a-b) and target injections to V2 (Fig. 1c-73 d) (see Methods). Linear array recordings were, subsequently, targeted to GFP/ArchT-expressing 74 V1 regions (Fig. 1c,e). Trial interleaved and balanced surface laser stimulation of increasing 75 intensity was applied to ArchT-expressing axon terminals of V2 feedback neurons at the V1 76 recording site ( Fig. 1c; see Methods). This viral injection protocol produces ArchT-GFP 77 expression in V2 neurons at the injected site, including neurons sending feedback projections to 78 V1 but also other V2 neurons projecting within V2 itself or to other brain regions. However, 79 5 2.26±0.35°; infragranular layers 1.29±0.25º vs. 1.88±0.26º; T-test p<0.05 for all layers;  Whitney U-test, p < 0.05 for all layers). This suggests that, at least in the granular layer, which 125 does not receive direct feedback terminations, changes in RF size are relayed via other layers. 126 Since RF size derived from the empirically measured curves can be subject to noise, we 127 also compared the RF size with and without laser extracted from phenomenological model fits to 128 the summation data, as these can provide more robust measures of RF size. To this purpose, we 129 fitted to the summation data two different models, namely a ratio or difference of integrals of 130 two Gaussians (ROG or DOG model, respectively; see Methods), as these models have 131 previously been shown to provide a good description of spatial summation curves in macaque 132 V1 14,15 . In these models, a center excitatory Gaussian, corresponding to the RF center, overlaps a 133 spatially broader inhibitory Gaussian, representing the suppressive surround (see inset in Fig.  134 2d); the major difference between the two models is that the surround inhibits the center through 135 division in the ROG model, but through subtraction in the DOG model (see Methods). The ROG 136 model provided a better fit for most (79%), but not all, of the cells (see below). Therefore, we 137 fitted both models to the spatial summation data with and without laser stimulation, and for each 138 cell we extracted RF size from the model that provided the best fit to that cell's data. From the 139 fitted curve, RF size was defined as the stimulus diameter at 95% of peak response (as in 14 ) ( Fig.  140 2d inset). Importantly, we still found feedback inactivation to significantly increase RF size 141 when the latter was estimated from the models fits ( Fig. 2d However, responses to stimuli extending into the surround were increased in some cells ( Fig.  202 2a), but decreased in other cells (Fig. 2c inset). We asked whether different levels of laser 203 intensity had different impact on V1 neurons' response gain. 204 Figure 4a-b shows two example cells in which RF size progressively increased and 205 response amplitude progressively decreased with increasing laser intensity. However, the cell in 206 Figure 4b showed greater and overall response reduction, while for the cell in Figure 4a  207 response reduction was more pronounced at smaller stimulus diameters. Across the population of 208 cells (n=33) we found that 36% of neurons showed response reduction across the entire spatial 209 summation curve, and these were the neurons in the population that showed strongest surround 210 suppression in the no-laser condition (SI: 0.78±0.03.1% vs. 0.49±0.07%, T-test p<0.05). 211 7 We quantified how RF diameter and mean response amplitude varied with laser intensity. 212 This analysis is based on a population of 14 cells for which at least two laser intensities (within 213 the range selected on the basis of the control experiments described in Supplementary Figs.

2-214
3) induced significance changes in the spatial summation curve (ANOVA p<0.05); for each of 215 these cells the analysis was performed at the lowest (range: 3-31mW/mm 2 ) and highest (range: 216 18-43mW/mm 2 ) intensity. 217 Compared to lower laser intensity, at higher laser intensity 11/14 cells showed a 218 significant reduction in mean response amplitude (T-test p<0.05; Fig. 4c) and 10/14 cells showed 219 increased RF diameter (Fig. 4d). Furthermore, most cells (10/14) showed both, reduced response 220 gain and increased RF size with increasing laser intensity (Fig. 4e). For the cells that showed a 221 statistically significant gain change at higher laser intensity (n=11; black dots in Fig. 4e) summation data presented in Figure 2 in the laser and no-laser conditions, and compared how 232 well each model fitted the data (see Methods). We found that for the majority of the cells (79%) 233 the ROG model provided a better fit to the data (mean R 2 ±s.e.m. for cells that were best fit by the 234 ROG model 0.67±0.04 vs. 0.37±0.10 for the DOG model fits to the same cells). For the reminder 235 of cells (21%), both models provided similar good fits to the data. This result is consistent with 236 the idea that the surround affects neural responses with and without feedback via divisive 237 normalization mechanisms 14 . 238 We next determined which model parameters were mostly affected by feedback 239 inactivation. To achieve this goal, we selected for each cell the model that provided the best fit to 240 its size tuning measurements in the no-laser condition, and then allowed one parameter at a time 241 to vary with feedback inactivation, while the remainder of the model parameters were held fixed 242 to their original values. As none of the single parameter models could account for the full range 243 of the effects seen in the inactivation data (see Supplementary Results and Supplementary Fig.  244 5a), we next allowed two parameters at a time to vary with feedback inactivation, while holding 245 the rest constant, and performed this analysis for all possible combinations of parameter pairs. 246 The model in which both the spatial extent and gain of the center excitatory mechanism were 247 allowed to vary best accounted for the inactivation results of 30% of cells in the population, 248 followed by a model in which the spatial extent of both the excitatory center and inhibitory 249 surround mechanisms were varied, which, instead, provided best fits for 21% of the cells (see 250 Supplementary Results and Supplementary Fig. 5b-c). However, none of the two-parameter 251 models provided best fit for the majority of the cells. Moreover, when comparing the different 252 models based on the coefficient of determination (R 2 ) distributions, rather than fraction of cells 253 best fit by each model, we found that the different models performed similarly (see 254 Supplementary Results and Supplementary Fig. 5d). 255 8 To gain better insights into the circuit mechanisms underlying changes in RF size and 256 response gain induced by feedback inactivation, we used neural network modeling. As the 257 effects of feedback inactivation on spatial summation are reminiscent of the effects of changing 258 stimulus contrast 14,36 , we used a previously published recurrent network model of V1, which well 259 described the spatial summation properties of V1 cells and their contrast dependence. In 260 particular, we asked whether similar circuit mechanisms underlying changes in RF size and gain 261 with stimulus contrast could also account for the effects of feedback inactivation. 262 We used the 1D recurrent network model of Schwabe et al. 37 , which accounts for 263 surround suppression in V1 using intra-V1 horizontal and local recurrent connections, feedback 264 connections from a single extrastriate area, and a single population of inhibitory (I) neurons ( Fig.  265 5a; see Methods). In this model, I neurons have higher threshold and gain than excitatory (E) 266 neurons ( Fig. 5b) and, consistent with recent findings 38 , are more strongly driven by horizontal 267 connections than the E cells whose output they control. As a result, I cells generate suppression 268 under sufficiently high levels of excitation, but are inactive for low levels of excitation. 269 Therefore, the local network in the model becomes more dominated by inhibition with increasing 270 excitatory drive. For weak excitatory inputs (e.g. small visual stimuli in the RF), I neurons are 271 silent, but for strong inputs (e.g. large stimuli encompassing the RF and surround), they become 272 active ( Fig. 5c dashed pink curve) and suppress the E neurons' response ( Fig. 5c  intensity, leads to both further increase in RF size and further decrease in response gain ( Fig. 5d  284 solid green curve). This is consistent with the behavior of most cells in Fig. 4e (data points in the 285 shaded squares), for which we indeed found a significant negative correlation between RF size 286 change and response gain change when laser intensity was increased. Therefore, a single 287 mechanism in the network model can account for the main effects of feedback inactivation, i.e. 288 increased RF size and response gain change. 289 The network model could not easily reproduce the overall strong reduction in response 290 amplitude of the entire summation curve, as seen in 36% of cells, particularly at higher laser 291 intensity (e.g. Fig. 4b), perhaps because it relies on a single inhibitory neuron type. Moreover, 292 V1 receives feedback connections from multiple extrastriate areas, whose spatial extent increases 293 with the area's hierarchical distance from V1 34 . As the model incorporates feedback connections 294 at a single spatial scale, it cannot optimally reproduce the differential effects on proximal vs. 295 distal surround suppression of removing feedback from a single area, while leaving intact more 296 extensive feedback from other areas. Specifically, far surround suppression in the model was 297 weaker than in the data. Thus, future refinements of this model will have to incorporate feedback 298 at multiple spatial scales and multiple inhibitory neuron types. 299 300 9 301 DISCUSSION 302 303 Our study elucidates how feedback affects neural responses in the primate early visual cortex. 304 Reducing V2 feedback activity increased RF size, decreased V1 cell's responses to stimuli 305 confined to their RF, and increased their responses to stimuli extending into the proximal 306 surround, thus weakening surround suppression. The magnitude of these effects depended on the 307 degree of feedback inactivation, so that stronger reduction of V2 feedback activity led to greater 308 increase in RF size and progressive decrease in response amplitude. Therefore, our results 309 indicate that feedback from V2 controls RF size, proximal surround suppression and response 310 gain in V1. 311 Our study is the first to demonstrate that feedback is part of the network that regulates the 312 RF size of V1 neurons. None of the previous studies reported systematic effects of inactivating 313 extrastriate cortex on V1 cells' RF size [23][24][25][26][27][28][29] . For most of these previous studies, this is because 314 RF size was not measured after inactivation of higher cortical areas 23,24,26,27,29 . In two prior 315 studies 25,28 , however, spatial summation measurements similar to those performed in our study 316 were made before and after inactivation of higher visual cortex. It is unclear why no systematic 317 effects of inactivating extrastriate cortex on RF size were observed in these two studies, but 318 differences with our study that could have led to the different results include the specific cortical 319 areas that were inactivated (macaque V2 and V3 25 , or cat postero-temporal visual cortex 28 , likely 320 homologue of macaque inferotemporal cortex), inactivation methods (cooling of entire cortical 321 area/s), and data analysis. Compared to previous studies, which silenced an entire cortical area, 322 therefore also affecting activity in downstream cortical or subcortical areas, the strength of our 323 approach is the selective and titrated manipulation of feedback neuron activity. including levels and spatial extent of feedback inactivation, the specific cortical area inactivated 360 (two of these studies inactivated higher level cortical areas), and methods of quantifying 361 surround suppression that did not take into account the spatial extent of the specific feedback 362 system that was inactivated. 363 Inactivating V2 feedback reduced suppression predominantly in the proximal surround, 364 and did not abolish distal surround suppression. This was predicted on the basis of the known 365 visuotopic extent of V2 feedback connections. The latter are less extensive than feedback 366 connections arising from areas V3 and MT, which, instead, encompass the full extent of the 367 distal surround 34 . What may be surprising is that inactivating V2 feedback had no effects on 368 distal surround suppression. Linear summation predicts a reduction (but not abolishment) of 369 suppression caused by the largest stimuli when V2 feedback inactivation leads to reduced 370 proximal surround suppression. Therefore, our finding suggests that feedback from different 371 extra-striate areas affect V1 responses via a common non-linear mechanism. 372 To gain insights into the mechanisms underlying the impact of feedback on V1 neuron 373 responses, we fitted the data with phenomenological models previously used to describe the 374 effects of contrast on RF size 14,36 , as well as the effects of inactivating areas V2 and V3 on 375 surround suppression in V1 47 . In these models, the RF and surround have Gaussian sensitivity 376 profiles, with the RF described as an excitatory Gaussian and the surround as an inhibitory 377 Gaussian, the two interacting either subtractively or divisively. Sceniak et al. 36 found that at low 378 stimulus contrast, RF size is larger and response gain is lower than at high contrast, and 379 suggested this results from an increase in the spatial extent of the center Gaussian mechanism. 380 Cavanaugh et al. 14 , instead, demonstrated that contrast-dependent changes in RF size and gain 381 could be explained by changes in the gain of both the center and surround Gaussian mechanisms. 382 Our modeling results differ from these previous reports, because although the effects of contrast 383 on RF size and gain resemble some of the effects of feedback inactivation, particularly those we 384 have observed at higher laser intensity, they nevertheless represent only a subset of the full range 385 of feedback inactivation effects. Thus, models in which feedback inactivation modifies only the 386 spatial extent of the center Gaussian 36 , or only the gain of both the center and surround 387 Gaussians 14 could capture the increase in RF size and gain reduction, but failed to capture the 388 simultaneous response decrease to stimuli in the RF and response increase to stimuli extending 389 into the proximal surround. 390

11
The modeling work of Nassi et al. 47 showed that changes in the spatial extent of 391 inhibition best accounted for changes in V1 spatial summation after simultaneous cooling of 392 macaque areas V2 and V3. In agreement with this previous study, we found that such a model 393 could capture the changes in neural responses for stimuli extending into the surround, i.e. the 394 reduction in surround suppression found in both our and these authors' study. However, in 395 contrast to Nassi et al., we found that feedback inactivation also caused an increase in RF size, 396 and this could not be accounted for by a model in which feedback only affects the spatial extent 397 of surround inhibition. Instead, we found that a model involving changes in the spatial extent and 398 gain of the excitatory mechanism best accounted for the range of feedback inactivation effects, 399 suggesting that V2 feedback affects the spatial extent over which cells integrate excitation as 400 well as the gain of excitation. 401 A simple network model in which spatial summation results from the interaction of 402 feedforward, V1 horizontal and inter-areal feedback connections with local recurrent networks, 403 provided further insights into the network mechanisms that may underlie these effects of 404 feedback inactivation. In this model, changes in RF size and response amplitude after feedback 405 inactivation were explained by a single mechanism, asymmetric inhibition, which leads to an 406 altered balance of excitation and inhibition when excitatory feedback inputs to E and I neurons 407 are reduced. While in our model asymmetric inhibition is implemented using high-threshold/gain 408 somatostatin-like inhibitory neurons, in principle other models with asymmetric inhibition 409 should be able to account for feedback inactivation effects on RF size and response amplitude. 410 For example, in the model of Rubin et al. 48 , asymmetric inhibitory/excitatory responses are 411 implemented using a mechanism based on a supralinear input/output function of cortical neurons 412 (which causes the gain of the input/output function to increase with increasing postsynaptic 413 activity) and an inhibition-stabilized network (in which strong recurrent excitation is stabilized 414 by strong recurrent inhibition (1-2%), and end-tidal CO2, blood oxygenation level, electrocardiogram, and body temperature 442 were monitored continuously. The scalp was opened and the skull was thinned using a dental 443 drill over areas V1/V2, covered with agar and a coverslip, which was glued to the skull. On 444 completion of surgery, isofluorane was turned off, anesthesia maintained with sufentanil citrate 445 (8-13μg/kg/hr, i.v.), and paralysis was induced with repeated 30-60 min intravenous boluses of 446 rocuronium bromide (0.6mg/kg/hr) to stabilize the eyes. The pupils were dilated with a topical 447 short-acting mydriatic agent (tropicamide), the corneas protected with gas-permeable contact 448 lenses, the eyes were refracted, and optical imaging was started. Once the V1/V2 border was 449 functionally identified, the glass coverslip was removed, small craniotomies and durotomies 450 were performed over V2, and the viral mixture slowly pressure-injected (240nl/site at 500µm 451 and again at 1200µm depth, using glass pipettes of 40-50μm tip diameter, 15 minutes/240nl). 452 The thinned skull was reinforced with dental cement, the skin sutured and the animal recovered. 453 454 Optical Imaging 455 Acquisition of intrinsic signals was performed using the Imager 3001 (Optical Imaging Ltd, 456 Israel) under red light illumination (630 nm). Imaging for orientation and retinotopy allows 457 identification of the V1/V2 border ( Fig. 1a-b). Orientation maps were obtained using full-field, 458 high-contrast (100%), pseudorandomized achromatic drifting square-wave gratings of 8 459 orientations at 0.5-2.0 cycles/° spatial frequency and 2.85 cycles/s temporal frequency, moving 460 back and forth, orthogonal to the grating orientation. Responses to same orientations were 461 averaged across trials, baseline subtracted, and difference images obtained by subtracting the 462 response to two orthogonal oriented pairs (e.g. Fig. 1b middle panel). Retinotopic maps were 463 obtained by subtracting responses to monocularly presented oriented gratings occupying 464 complementary adjacent strips of visual space, i.e. masked by 0.5-1° strips of gray repeating 465 every 1-2°, with the masks reversing in position in alternate trials (Fig. 1b right panel) 51 . In each 466 case, reference images of the surface vasculature were taken under 546 nm illumination (green 467 light, Fig. 1b left panel), and later used as reference to position pipettes for viral vector injection. 468 469

Electrophysiological Recordings and Visual Stimulation 470
Following 62-68 days after the viral vector injection, animals were anesthetized and paralyzed by 471 continuous infusion of sufentanil citrate (6-13µg/kg/h) and vecuronium bromide (0.3mg/kg/h), 472 13 respectively, and vital signs were continuously monitored, as described above. The pupils were 473 dilated with topical atropine, protected with lenses and refracted. GFP-expressing V2 injection 474 sites and V2 feedback axonal fields in V1 were identified with GFP goggles (Fig. 1e1), and small 475 craniotomies were made over V1. Extra-cellular recordings were made in V1 with 24-channel 476 linear multielectrode arrays (V-Probe, Plexon, Dallas, TX; 100μm contact spacing, 20μm contact 477 diameter) coated with DiI (Molecular Probes, Eugene, OR) to assist with post-mortem 478 reconstruction of the electrode penetrations (e.g. Fig. 1e2), and lowered normal to the cortical 479 surface (using triangulation methods) to a 2-2.2 mm depth over 60-90min. A 128-channel system 480 (Cerebus, Blackrock Microsystems, Salt Lake City, UT) was used for signal amplification and 481 digitization (30 kHz). Continuous voltage traces were band-pass filtered (0.5-14.25 kHz), and 482 spikes were detected as spatiotemporal waveforms using the double-threshold flood fill 483 algorithm 52 (thresholds 2 and 4 x noise S.D.). This procedure was adopted because the apical 484 dendrites of pyramidal cells run parallel to the probe shank and may spread the same waveforms 485 across multiple channels. A masked EM algorithm 53 was used for clustering, and manual 486 refinement of the clusters was performed with the Klustasuite 52 . 487 After manually locating the recorded RFs, their aggregate minimum response field was 488 quantitatively determined using a sparse noise stimulus (500ms, 0.0625-0.25 deg 2 square, 489 luminance decrement, 5-15 trials; Supplementary Fig. 1b) and all subsequent stimuli were 490 centered on this field. Orientation, eye, spatial and temporal frequency preferences for the cells 491 in the recorded V1 column were determined using 1º diameter, 100% contrast drifting sinusoidal 492 gratings monocularly presented on an unmodulated gray background of 45cd m -2 mean 493 luminance. We then performed spatial summation measurements using circular patches of 100% 494 contrast drifting sinusoidal gratings of increasing diameter centered over the columnar aggregate 495 minimum response field. The patch diameter ranged from 0.2-0.6° to 10-18° (depending on 496 animal) and different patch sizes were presented in random order within each block of trials. All 497 size-tuning experiments were performed using gratings of spatial and temporal frequencies and 498 orientation that strongly drove most cells in the column. It was not typically challenging to find 499 spatial and temporal frequency values to which all cells in the column responded vigorously. 500 When the penetration was perfectly vertical, orientation preference was also similar for all cells 501 in the column. Slight deviations from vertical, however, even for RFs perfectly aligned in space, 502 could cause orientation to shift slightly across the column, due to the narrow orientation tuning 503 of many V1 cells 54 . In this case, the size tuning experiment was run using two different 504 orientations. Importantly, although deviations from optimal stimulus parameters can increase the 505 neurons' summation area 55 , these deviations are not expected to cause differences between 506 neuronal responses recorded with and without laser stimulation. To monitor eye movements, the 507 RFs were remapped by hand approximately every 10 minutes, and stimuli were re-centered in 508 the RF when necessary. Stimuli were presented for 500ms with 750ms inter-stimulus interval. 509 Stimuli were programmed with Matlab (Mathworks, Natick, MA) and presented on a linearized 510 CRT monitor (Sony GDM-C520, 600 x 800 pixels, 100Hz, 57cm viewing distance) and their 511 timing was controlled with the ViSaGe system (Cambridge Research Systems, Cambridge, UK). 512 Data analysis was performed using custom scripts written in Matlab and Python 56,57 . 513 514 Laser Stimulation 515 14 optical fiber, then expanded and collimated to a 2.8 mm spot. Reported irradiances refer to the 517 light power exiting the collimator divided by the area of the collimator. Because the beam was 518 collimated, the illumination spot size depended very little on the distance of the fiber from the 519 brain. Laser timing was controlled at submillisecond precision, using custom made programs 520 running on real-time Linux. Light was shone on the surface of V1 through thinned skull in the 521 regions of GFP expression, and V2 was shielded from light. Laser onset was simultaneous with 522 stimulus onset and photostimulation continued throughout stimulus presentation (500ms). The 523 animal's eyes were shielded from the laser light. 524 525 Neuronal Sample Selection 526 We analyzed 67 visually responsive (defined as max response at least 2SD>baseline) and 527 stimulus modulated (one-way ANOVA, p<0.05) units. Approximately 61% (41/67) of these 528 visually driven single-units were modulated by one or more laser stimulation intensities (two-529 way ANOVA, either laser or stimulus diameter x laser interaction, p<0.05, or at least two 530 successive data points different in the same direction, p<0.05). We were not able to determine 531 RF size for eight cells, thus these were excluded from further analysis. Therefore, a total of 33 532 cells were analyzed for the results reported in Figs unit separately, and the analysis was performed at this intensity. This was motivated by 539 expectations that the light intensity required to produce inactivation effects differs among cells 540 due to several factors, including variation in opsin expression across neurons, distance of the 541 cells from the light source, and intrinsic differences in sensitivity to feedback perturbation. 542 Importantly, however, even though we selected different light intensities for different cells, the 543 direction of the effects was not biased by our analysis, as we selected for each cell the laser 544 intensity causing the largest change in RF size, irrespective of whether this was an increase or 545 decrease. 546 The analysis of the data presented in Fig. 4 is based on a population of 14 cells for which 547 at least two laser intensities (within the range selected on the basis of the control experiments 548 described in Supplementary Figs. 2-3) induced significance changes in the spatial summation 549 curve (ANOVA for either laser or stimulus diameter x laser interaction p<0.05). This is a subset 550 (14/33) of the population analyzed in Figs. 2-3, because for the remainder of the population we 551 either lacked two laser intensity levels, or only one laser intensity (within the range selected on 552 the basis of control experiments) caused significant changes. 553 554 Definition of RF and Surround Size 555 From the size tuning curves, measured as described above, for each cell we extracted as a 556 measure of RF size the grating's diameter eliciting maximum response. Surround size was 557 defined as the smallest grating diameter for which the neuron's response was reduced to within 558 5% of the response at the largest diameter. As these measures of RF and surround size can be 559 subject to noise, to derive more robust measures, we also fitted the size tuning data with the ratio 560 and difference of the integral of two Gaussian functions (ROG 14 and DOG 15 models,561 respectively; see below for model fits). From the fitted summation curves we extracted the cells' 562 RF size as the smallest stimulus diameter at which the cell response reached 95% of the peak 563 response 14 . 564 565 Statistical Model Fitting 566 ROG 14 (eq. 1) and DOG 15 (eq. 2) models were fitted to the size tuning data according to the Here the variable x corresponds to the diameter of the stimulus, wc and ws are the spatial extents 580 of the center excitatory and surround inhibitory Gaussian mechanisms, respectively (with the 581 constraint that wc < ws), Lc and Ls are the activities of the center and surround mechanisms, 582 respectively, and gc and gs are the gains of the center and surround mechanisms, respectively. All 583 parameters were constrained to positive values during optimization. Model parameters were 584 optimized by minimizing the sum of squared errors between the model predictions and the data. 585 Initial parameter search was done by performing two successive grid optimizations. The first grid 586 was coarse, and the second grid was finely spaced and centered on the best fitting parameters 587 determined with the first grid search. The best fitting parameters determined with the second grid 588 were used as initial parameters for final optimization, which was done using the active-set 589 algorithm in Matlab. As the models have an equal number of parameters, model comparisons 590 were performed by directly comparing coefficient of determination (R 2 ) values. R 2 values were 591 estimated using linear regression. 592 593

Identification of Laminar Borders and Analysis of RF Alignment 594
To ensure that the array was positioned orthogonal to the cortical surface, we used as criteria the 595 vertical alignment of the mapped RF at each contact (see Supplementary Fig. 1b), as well as the 596 similarity in the orientation tuning curves recorded at each contact. The array was removed from 597 cortex and repositioned, if significant RF misalignments across contacts were detected. The 598 degree of RF misalignment was also quantified for each penetration as described in the 599 Supplementary Results (see Analysis of RF Alignment). 600 The borders between the granular layer (4C) and supra-and infragranular layers were 601 determined by applying current source density (CSD) analysis, using the kernel CSD method 58 , 602 to the band-pass filtered (1-100 Hz) and trial averaged (n=400) continuous voltage traces evoked 603 by a brief full-field luminance increment (100ms, every 400ms, 1-89cd m -2 ; Supplementary 604 Fig. 1a). As previously established 59 , the earliest current sink corresponds to the granular layer, 605 and its borders with the supra-and infra-granular layers can be determined from the reversals 606 from current sink to current source above and below the granular layer, respectively. 607 608 Statistical Analysis 609 Statistical p-values refer to either independent sample or one sample two-tailed T-tests. For the 610 within layer comparisons (Fig. 2b4), where the expected effect direction was known, one-tailed t-611 tests are reported. When deviations from normality were detected using QQ-plots (RF size 612 analysis), the T-tests were augmented with Mann-Whitney U-test. The Suppression Index (SI) in Fig. 3a2-a3 was computed as follows: SIno-laser= (RC-no-laser -RCS-no-619 laser)/RC-no-laser. SIlaser= (RC-no-laser -RCS-laser)/RC-no-laser, where RC-no-laser is the response to a 620 stimulus confined to the RF (the peak of the summation curve) in the no-laser condition, RCS-no-621 laser is the response to the stimulus covering the RF and surround in the no-laser condition (the 622 proximal surround only for the measurements in Fig. 3a2, and the full extent of the surround for 623 the measurements in Fig. 3a3), and RCS-laser is the response to the stimulus covering the RF and 624 surround in the laser condition. 625 626 Histology 627 On completion of the recording session, the animal was perfused transcardially with 2-4% 628 paraformaldehyde in 0.1M phosphate buffer. The occipital pole was frozen-sectioned at 40μm, 629 tangentially to the cortical surface (n=2 brains), or sagittally (n=1). GFP label in V2 and V1 and 630 DiI tracks were visualized under fluorescence to ascertain injection sites were confined to V2, 631 and electrode penetrations were targeted to regions expressing GFP (Fig. 1d,e2). Electrode 632 penetrations from regions with low GFP expression were eliminated from analysis. Sections 633 were counterstained with DAPI (Sigma-Aldrich, St. Louis, MO) to identify V1/V2 border and 634 cortical layers (Fig. 1d2).

636
Network Model 637 The network mechanisms underlying the observed effects of feedback inactivation were 638 investigated using the model of Schwabe et al 37 . We used exactly the same recurrent network 639 architecture and parameters as in the original published model, which was shown to capture 640 several response properties of surround suppression in V1, including contrast-dependent changes 641 in RF size and surround suppression strength. 642 For model details we refer the reader to the original publication. Briefly, the network 643 model represents two areas of visual cortex, V1 and an extra-striate area, each area simplified to 644 a single cortical layer. A schematic diagram illustrating the basic network architecture is shown 645 in Fig. 5a. Each spatial location in the model is represented by coupled local excitatory (E) and 646 inhibitory (I) cells, which act as the basic functional module of the network that incorporates the 647 effects of local recurrent connections. Interactions between these modules are mediated by 648 horizontal and feedback connections. The spatial profile and conduction velocities of horizontal 649 and feedback connections are constrained by existing anatomical and physiological data, 650 according to which feedback connections are spatially more extensive 34 and have faster 651 conduction velocities 45 than horizontal connections. Because we are focusing on size-tuning 652 effects in this study, it seemed sufficient to take a very simple local network model with a single 653 inhibitory neuron type. The stimulations were run with 30% contrast, which is equivalent to 654 translating the contrast response functions of the model neurons along the contrast axis. This 655 modification is justified as V1 neurons exhibit a variety of contrast preferences. 656 657 Data Availability 658 The data will be made available upon reasonable request to the authors. 659 (1999).   For each array penetration, we quantified the alignment of the minimum response fields 5 (mRFs) mapped at each contact across the array. For each contact, we calculated the 6 Euclidean distance of the mRF center, mapped at that contact, from the mean of mRF 7 centers in that penetration. For the purpose of this analysis, the mRF center was defined 8 as the center of mass of the thresholded mRF activity map (responses >1SD above 9 baseline; Supplementary Fig. 1b). The median Euclidean distance from the mean mRF 10 center of each penetration was 0.124º±0.029º/0.035º (95% CI lower bound/upper bound; 11 bootstrap). 12 13 14

Control Experiments in Cortex Not Expressing ArchT 15
For the main experiment, laser intensities were selected based on a control experiment in 16 one animal (n=2 penetration) in cortex not expressing ArchT. Recordings and analysis 17 were otherwise identical to the main experiment. Unsorted, thresholded multi-units were 18 analyzed for this control. 19 We found light artifacts at relatively low light intensities (63mW/mm 2 ; see 20 Supplementary Fig. 2a), which, to our surprise, have been used in previous optogenetic 21 experiments. The laser artifacts were qualitatively different in superficial and deep layers: 22 spike-rates were usually increased in superficial layers, but often decreased in deep layers 23 (Supplementary Fig. 2a). For granular and infragranular layers, irradiances at or below 24 43mW/mm 2 did not produce statistically significant changes in the cells' size tuning 25 curves ( Fig. 2a-b) Supplementary Fig. 2d). 41 2 The analysis reported in the Results (Figs. 2-3) was performed for laser intensities 42 up to 43mW/mm 2 , because the laser artifacts induced in some supragranular cells at this 43 intensity could not account for the observed effects of feedback inactivation (as these 44 artifacts caused decreases rather than increases in RF size). However, to further 45 corroborate that our results of feedback inactivation could not be attributed to laser-46 induced artifacts, we repeated all the main analyses of data recorded in ArchT-expressing 47 cortex, after excluding supragranular units which showed inactivation effects at laser 48 irradiances >19mW/mm 2 , i.e. irradiance levels that had produced artifacts in some 49 supragranular layer cells in control cortex. The results of this analysis were qualitatively 50 and quantitatively similar to those of the original analysis, as detailed below. 51 52 Analysis of Data in Cortex Expressing ArchT, Excluding Supragranular Cells Showing 53 Inactivation Effects at >19mW/mm 2 Irradiance. 54 Mean RF diameter was significantly smaller with intact feedback, compared to when 55 feedback was inactivated (mean±s.e.m no-laser vs. laser: 1.24±0.11º vs. 1.83±0.17º, 56 p=0.007, n=26; Supplementary Fig. 3a), with a mean increase of 59.3±13.0% 57 (p<0.001). As for the original analysis (Fig. 3a) Fig. 3c). We conclude that increased RF diameter and reduced surround 70 suppression indeed resulted from inactivating V2 feedback to V1, and were not caused by 71 laser-induced heat artifacts. 72 None of the units recorded in the control experiment showed reduced response at 73 the irradiances used for the analysis of data in Fig. 4. Thus, we are confident that the 74 general response suppression for small and large stimuli observed in the data reported in 75 Fig. 4 resulted from inactivating feedback axons. 76 77 78

Analysis of RF Size Increase Induced by Noise 79
We performed an analysis to exclude the possibility that the increased RF size after 80 feedback inactivation could arise from noise in the spatial summation data. For each cell, 81 we first generated a size-tuning curve from the fitted ROG or DOG model, depending on 82 which model better fitted the cell's response. For each cell, we then generated a "noisy" 83 curve by independently sampling for each presented stimulus diameter 10 responses from 84 a Poisson distribution having the same mean as the "real" fitted curve at those diameters. 85 These 10 sampled responses per stimulus diameter were then averaged to produce a noisy (a) The black curve represents the "real" size tuning curve derived from phenomenological model fits to the spatial summation data of an example V1 cell. The red circles represent the simulated response at each stimulus diameter averaged over 10 trials. The response in each trial was obtained by randomly sampling from a Poisson distribution with the same mean as in the "real" curve. (b) Cumulative distribution of percent RF size changes expected under the null hypothesis that the real size tuning curve measured with and without laser stimulation were identical, and all changes in RF size were due to noise. RF size change due to noise was computed separately for each cell, based on a "noisy" tuning curve formed by averaging simulated responses in 10 trials, then averaging over the population of 33 cells and repeating the procedure 10,000 times (see Supplementary Results for details).