Novel Object Detection and Multiplexed Motion Representation in Retinal Bipolar Cells

Antagonistic interactions between the center and surround receptive field (RF) components lie at the heart of the computations performed in the visual system. Center-surround RFs are thought to enhance responses to spatial contrasts (i.e., edges), but how they contribute to motion processing is unknown. Here, we addressed this question in retinal bipolar cells, the first visual neuron with classic center-surround interactions. We found that bipolar glutamate release emphasizes objects that emerge in the RF; their responses to continuous motion are smaller, slower, and cannot be predicted by signals elicited by stationary stimuli. The alteration in signal dynamics induced by novel objects dwarfs the enhancement of spatial edges and can be explained by priming of RF surround during continuous motion. These findings echo the salience of human visual perception and demonstrate an unappreciated capacity of the center-surround architecture to facilitate novel object detection and multiplexed encoding of distinct sensory modalities.


Main text: Introduction
The ability to detect motion begins in the retina, which contains ganglion cells dedicated to the detection of local motion 1-3 , approaching objects 4 , acceleration 5 and the direction of movement (for a review, see 6,7 ). The highly specialized computations in ganglion cells are driven 25 and shaped by glutamate release from axonal terminals of bipolar cells (BCs), which in mice are divided into about 14-15 functional types that are tuned to different visual features 8- 10 . The topographic stratification of BC axons in the inner plexiform layer (IPL) establishes some of the functional organization of visual processing in the retina: BCs that carry ON signals (depolarization to light) are found closer to the ganglion cell layer, and cells with sustained 30 responses are segregated towards IPL borders 8, [11][12][13] . The difference in visual processing between BCs reflects their center-surround architecture, comprised of two separate concentric regions sampling the visual signal. This RF structure is formed by direct innervation of BC dendrites by photoreceptors in their excitatory center and a combination of horizontal and amacrine cell inhibition in the antagonistic surround 9,14-16 . 35 Historically, motion signals in BCs have been understood as a linear combination of static responses, much like how the perception of motion is produced in movies by a rapid presentation of discrete images 17,18 . However, computations in cells with centre-surround RFs are inherently nonlinear [19][20][21][22] and depend on the spatiotemporal RF activation pattern, which differs between moving and static stimuli. Thus, despite the abundance of the classic centre- 40 surround RFs in the early visual system, little is known about their impact on motion processing in general and on BC activity in particular.
To examine the properties of visual processing of moving objects in BCs, we recorded the change in glutamate levels across different depths of the inner plexiform layer (IPL) and captured the release dynamics of different BC types to moving or stationary bars. We reveal 45 significant alteration in the peak and the temporal characteristics of the glutamate responses following object motion. Additionally, our results indicate that BCs can signal the appearance of novel objects that enter the visual scene. Flashed stationary objects or stimuli that emerge from behind static occluders provoke intense discharge from all BCs, whereas continuous motion and disappearing stimuli suppress BC activation. These observations were not affected by the 50 pharmacological blockage of amacrine cell inhibition. Accordingly, a detailed simulation of signaling in the outer retina replicates the diversity of motion responses in BCs and reveals how motion computations can be carried out at the first retinal synapse by a horizontal cell-derived inhibitory signal and influence the representation of a realistic visual input. Our results reveal a fundamental property of signal integration in center-surround RFs to identify newly appearing 55 visual stimuli and diversify the representation of static and moving shapes.

Glutamate responses in BCs to full-field motion are diverse and do not follow the response dynamics for stationary signals
To study the representation of moving stimuli in BCs, we used two-photon microscopy 60 to collect light-driven glutamatergic signals in whole-mount mouse retinas expressing iGluSnFR, responding to static flashes and full-field moving bars 9,13,23 . We systematically surveyed all layers of the inner plexiform layer (IPL) with multiple scan fields; pixels with similar responses were then grouped into regions of interest (ROIs, Figs. 1a-b, s1, s2). The spatial extent of most ROIs was smaller than 50 µm, indicating sampling from a single cell or at most two functionally 65 similar BCs (Figs. 1b, s1) 9,24 . Responses to stationary flashes were used to combine ROIs from different experiments into functional clusters 9,11,24 . The optimal separation was obtained with 5 OFF and 7 ON clusters (Figs. 1c-d); comparable to previous classifications of glutamate signals in the IPL 9,11 . Following clustering, we analyzed responses to moving bars. As expected, slower RF engagement prolonged motion response kinetics ( Fig. 1d-f). Surprisingly, there was no 70 correlation between the flash-and motion-driven rise time dynamics (Fig. 1f, left, Pearson correlation coefficient, R=-0.04).
It is possible that the difference in motion processing we describe reflects the topographic stratification of BC axons with sustained responses approximating the IPL borders 8, [11][12][13] . To assess this, we analyzed signal parameters relative to recording depth (Fig. 1g) 75 or signal transiency index (TI, calculated from stationary response kinetics, Fig. 1g-h). We identified a clear relationship between cluster transiency to the change in the amplitude (R = -0.95) and the decay-time (R = 0.87) of motion responses relative to the stationary signals (Fig.   1h). Similarly, the effect of motion on the peak amplitude and decay time was greatest in the central regions of the IPL, reflecting the stratification level of the transient BCs (Fig. 1g). In 80 contrast, the change in the rise-time did not follow the transient-sustained division (Fig. 1h).
Instead, we observed a gradual decrease in the motion/stationary ratio for the rise-time kinetics with increasing depth in the retina (Fig. 1g). Overall, the observed low correlation in key aspects of response shape and the distinct pattern of signal dependency on IPL depth between static and moving objects indicate a multiplexed representation of motion and stationary 85 information in the BC population.
Notably, these observations are not an artifact of our clustering approach, as our algorithm was agnostic to motion information. We conducted several tests to rule out the possibility that the results we describe here are due to the grouping of pixels with different recruitment times during motion responses. First, at odds with the predicted effects of such 90 pixel averaging, the degree to which motion impacted signal dynamics varied systematically between clusters, and the inter-cluster variability of responses was higher during motion ( Fig.   1d). Second, the mean responses recorded for each group closely mirrored the signals recorded in individual pixels (Fig. s2). Last, neighboring regions of the retina respond sequentially to motion, and for this reason, the influence of pixel averaging should be most evident in groups 95 with wide spatial pixel distribution. In contrast to this prediction, however, we found that the spread of each group's pixels along the axis of motion was not correlated with the response dynamics (Fig. s3).

Enhanced representation of novel stimuli
Previous work demonstrated that neurons could employ a simple strategy of comparing 100 the spatial extent of center-surround recruitment to detect local spatial contrasts 19,25,26 and diversify the representation of flashed objects 9,21 . According to the classic description of the center-surround interactions, occluders masking part of the surround enhance RF output ( Fig.   2a 'Edge'). We reasoned that responses to moving stimuli should also be sensitive to stationary edges in the RF. To explore this possibility, we presented horizontally moving bars and masked 105 the stimulus on the left or the right halves of the display.
Unexpectedly, the kinetics and the amplitude of the glutamate release were significantly faster/higher for bars emerging from the mask than for motion in the opposite direction ( Fig. 2 'Emergence' vs. 'Exit'). Across all ROIs, the peak response amplitude following emerging motion was significantly higher than the signal observed during full-field motion (122±2% mean±SEM; 110 p<0.001 vs. full-field motion, ANOVA followed by Tukey test), and the rise-time was sharpened by more than 50% (Fig. 2d). In comparison, the mean(±SEM) ratio between responses to fullfield flashes and motion in the same ROIs was 136±2.4% (p<0.001, ANOVA followed by Tukey test, n=365, Fig. 2). Thus, in terms of shape peak and temporal dynamics, the representation of emerging motion more closely resembles static flashes than continuous motion (Fig. 2). 115 We next separated between and compared transient (n = 182) and sustained (n = 183) ROIs to assess the effect of signal kinetics on visual processing in the presence of edges. The Compared to the classic role of center-surround in detecting spatial boundaries, we note that even though the occluding mask (when present) was identical for all protocols, we did not observe a significant effect of the masks on stationary responses (Figs. 2, s4 p>0.3 for peak/kinetics, ANOVA). We interpret this finding to indicate that RF structure in the early visual 125 system is tuned to highlight new information and not to detect inhomogeneous spatial compositions.

Amacrine cell inhibition is not required for novel object sensitivity and slower motion kinetics
What are the cellular components underlying motion computations in BCs? Previous work suggested that amacrine cell inhibition diversifies the representation of stationary stimuli 130 that partially occupy the RF of BCs 9 . Correspondingly, a cocktail of 50µM SR95531, 100µM TPMPA, and 1µM Strychnine to block GABAA, GABAC and glycine receptors (Fig. 3a) 9,24 changed the shape of glutamate waveforms elicited by stationary flashes (Figs. 3) and of calcium transients in ganglion cells (Fig. s5). Yet, motion signals in the IPL were not affected by the inhibitory blockers (Fig. 3). Because horizontal cells can control photoreceptor output by 135 mechanisms that do not require the release of neurotransmitters 16,27,28 , we reasoned that our pharmacological manipulation did not fully disrupt the horizontal feedback on the photoreceptors, suggesting that motion processing is performed already in the first retinal synapse. 140 To test whether signal interactions in the outer retina are sufficient to explain our experimental findings, we constructed a computational model of visual processing in the outer retina and BCs (Fig. 4). We activated the model with stationary and moving bars and recorded  (Figs. 4). In general, the encoding of existing objects is accompanied by a longer temporal delay between the initial activation of the surround and subsequent center stimulation. Due to this delay, surround inhibition is more developed by the time the center is engaged by the stimulus and is, therefore, more likely to suppress responses to continuous motion. 160 Focusing on the factors that influence BC response dynamics, we note that transient kinetics were mediated by faster neurotransmission, but also, unexpectedly, by elevated sensitivities to photoreceptor release (Figs. s8, s7). The higher threshold required for effectual activation increased the sensitivity of transient BCs to small fluctuations around the peak photoreceptor activity. In agreement with recent findings 22 , our model suggests that transient 165 BCs receive a more rectified, nonlinear copy of the photoreceptor signal and predicts that such nonlinearity creates a substrate for more distinct responses to motion vs. stationary stimuli and promotes the enhancement of novel object emergence (Figs. 4c-d, s9). 170 Next, we asked whether the fundamental properties of the center-surround RF architecture are sufficient to identify novel objects under realistic visual conditions. To address this question, we simulated responses from a population of linear center-surround neurons Although the simulated cells lacked nonlinear signal processing mechanisms, we found these cells 175 capable of generating a rich representation of dynamically changing scenes. Cells responding to established motion encoded the local contrast differences between the stimulus and the background (Figs. 5b-c). Comparable to our findings presented above, stimulus emergence correlated with robust responses (Fig. 5b-c, s10). Interestingly, novel motion enhancement was evident mainly at the initial site of stimulus appearance (the wing in the example shown in Fig.   180 5), implying a spatial focus for novel object detection spanning about 100 µm of retinal space.

Enhanced representation of novel stimuli under natural movies requires center-surround organization
Using the simulation, we were able to test the contribution of the surround to this computation. We reformulated the RF description for the tested population to exclude the surround. We found that the outputs of the cells in this simulation were still tuned to the local contrast ( Fig. 5d-e). However, the response amplitudes were similar for continuously moving 185 and emerging stimuli, indicating that similar to our findings in the simulated retinal circuit, novel object detection required surround participation (Fig. 5e, s10).
What is the benefit of utilizing the center-surround architecture to compute novel object appearance in a realistic environment? Stronger activation near the mask-stimulus boundary can be beneficial for detecting stimuli in downstream neurons. To quantify the 190 information that is encoded by individual neurons in our simulation, we measured the mutual information from responses of cells at the location of stimulus emergence. Analysis of signal entropies calculated from the peak responses to continuous and novel motion revealed that each cell is capable of transmitting 0.62±0.12 bits in each trial (Fig. 5f). Comparable information levels were found for responses in cells near vs. far (>200 µm) from the stimulus emergence

Edge effects influence the analysis of motion processing in the retina
Given the participation of BCs in novel motion detection, we asked whether the dependence of BC signals on the direction of motion near mask-stimulus boundaries impacts the computation of direction selectivity (DS). The earliest direction-selective signals are present in dendrites of starburst amacrine cells (SACs), which are tuned to detect stimulus motion 205 towards dendritic tips (Fig. 6a) [29][30][31][32][33] . Despite intense effort, explaining the biological implementation of this computation remains elusive 17,30,[34][35][36][37][38][39] . A common strategy to probe SAC DS is to isolate dendritic computations 6,40 with visual protocols structured to stimulate a part of the SAC 30,36,39,41 -effectively masking part of the stimulus (Fig. 6b). To explore whether the glutamatergic drive to SACs is affected by the mask-stimulus boundary, we expressed iGluSnFR 210 driven by the ChAT promoter ( Fig. 6a) 34,35 . We presented visual stimuli as above (Fig. 6b) and set the size of the field of view to match the span of BC innervation of a single SAC dendrite ( Fig. 6b, ~80 µm) 30,39 .
Akin to our other findings, glutamate responses were more pronounced for emerging stimuli (Fig. 6c-d). The mean (±SD) direction selectivity index (DSI) computed from moving bar 215 responses with the direction of motion towards/away from the boundary at the center of the display was 32 ± 21% (p<0.001 vs. 0, t-test, n=81 ROIs, Fig. 6e), while full-field moving stimuli evoked comparable glutamatergic responses in all directions (Fig. 6d-e). Could the directional effect observed in the presence of a mask-stimulus boundary contribute to DS computations in SACs? A simple analysis shows that the answer is no. The enhancement of BC drive aligns with 220 the preferred dendritic axis in SACs whose cell bodies happen to lie near the mask (Fig. 6b, e 'Edge near soma'). However, signals to SACs in less optimal configurations are in the 'wrong' direction. The grey-colored SAC illustrated in Fig. 6b serves as an example of a cell whose soma is located deeper in the stimulated region yet proximal enough to extend its dendrites over the mask ('Edge near tips'). With the direction of motion away from the mask-stimulus boundary 225 and towards the soma of this cell, stronger responses to emerging motion lead to a reversed directional tuning (Fig. 6e 'Edge near tips'), in contrast to what is expected of a proper directional mechanism.

Discussion
Using the retinal BCs as a model system, we were able to investigate the properties of 230 motion processing in center-surround RFs. We found that the representation of continuous motion was associated with reduced peak amplitudes and prolonged temporal dynamics of glutamate signals compared with sudden object appearance in most BC types. Motion responses could not be reliably predicted from the dynamics of responses to stationary flashes, indicating a multiplexed representation of static and moving objects (Fig. 1). Visual processing The unexpected diversity of motion responses revealed by our experiments highlights an asymmetric interaction between moving stimuli and static occluders; the encoding of object disappearance was similar to continuous motion, whereas newly emerging stimuli exhibited faster and more pronounced signals that qualitatively resembled the response to static flashes -250 particularly in transient BCs (Fig. 2).
Conceptually, our findings reflect a previously unappreciated capacity of centersurround RFs to signal the appearance of new objects. This property is a logical but previously undescribed consequence of the classic center-surround RF formulation. The mechanistic explanation for this function is straightforward and relies on the sequence of RF activation by 255 the stimulus. Continuously moving stimuli always enter the surround RF region first, priming the surround towards a more effective inhibition by the time the center is engaged. This process doesn't require any specific neuronal infrastructure and is present even in a linear RF formulation (Figs. 4, 5). Priming of the surround is weaker or absent in emerging motion and suddenly appearing objects. Correspondingly, the responses to these stimuli reflect the 260 stronger role of the center component in RF integration, leading to empirically observed enhanced response amplitudes and distinct temporal dynamics between the novel and existing objects.
In our hands, the computation of novel object detection was highly prominent across all BCs, whereas their spatial contrast sensitivity, as measured by the ratio between the responses 265 to static edges vs. full-field illumination, was not statistically significant (Fig. 2). These results suggest that in contrast to the prevailing view, sensitivity to spatial contrasts serves a secondary functional role in center-surround RFs, at least in the cells and the visual conditions we probed.
Our experiments and detailed circuit models show that photoreceptors and horizontal cells are the only circuit elements required to generate motion responses in BCs (Figs. 3, 4), 270 leading to the conclusion that the major steps in the computation of object motion already occur at the first synapse in the retina. Our findings support the idea that signal transformation from the photoreceptors to BCs could be nonlinear and that the degree of nonlinearity is larger for transient BCs 22 . Why is processing linearity correlated with the shape of the response? Our model of signal integration in the outer plexiform layer suggests a possible answer. Nonlinear 275 signal transformation at the photoreceptor-BC synapse could impose a threshold on the amplitude of the photoreceptor output that is required for effective activation of the postsynaptic cell (Fig. s7). As photoreceptors typically respond to light onset and light offset with a rapid membrane potential fluctuation 42,43 , nonlinear BCs are more likely to be disproportionally sensitive to these phases of photoreceptor release; their fast temporal 280 dynamics reflect the transient shape of the filtered photoreceptor output they sample.
Meanwhile, a linear signal processing mirrors the original shape of the photoreceptor light response (Figs. 4, s7). The exact biological implementation of the nonlinear photoreceptor-BC synapse dynamics is currently unclear but could plausibly be mediated by a differential affinity of BC dendrites to photoreceptor release 44 . In the end, the nonlinear nature of the transient BC 285 population is known to contribute to a rudimentary feature detector-like behavior that is tuned to certain visual conditions, such as signal polarity and spatial inhomogeneity 22 . We can now add novel object appearance to this list.
We used stimuli that were explicitly designed to compare responses to moving and static objects and representation of novel vs. existing visual items. Previous reports 290 demonstrated that the retina is capable of detecting acceleration 5 , differential motion 45 , looming (approaching) motion 4 , and distinguish between a motion to uncorrelated spatiotemporal activation 3 . These visual functions are thought to require higher-order retinal neurons (e.i., amacrine and ganglion cells). How RF components are integrated during the presentation of these stimuli and whether RF computations contribute to such sophisticated 295 calculations remain to be elucidated.
In the last decade, several groups found evidence for a spatial offset between presynaptic BC populations that are aligned with the directional axis in DS ganglion cells and dendritic position in SACs 17,24,30,46 . This circuit organization can support directional tuning by a mechanism first described by Hassenstein and Reichardt 47 -if the response speed of the BCs 300 follows their spatial arrangement. Conflicting results were reached in studies designed to test the predictions of this model using electrophysiological and imaging approaches 18,39,48 .
Importantly, all previous work examined BC output in response to the presentation of stationary inputs, which, as our results indicate, do not accurately reflect the dynamics in BCs during motion. Proposed directional computations are particularly dependent on BCs rise 305 times, which, as our data reveal, are uncorrelated between moving and static objects (Fig. 1).
At the very least, the dramatic increase in the rise-times dynamics we observed in our recordings suggests a shift in speed dependence of the Hassenstein-Reichardt detector to slowmoving objects. Further experiments will be required to resolve this issue and elucidate the potential impact of visual edges on observed DS (Fig. 6).

Imaging procedures
Mice were not dark-adapted to reduce rod-pathway activation. Two hours after enucleation, retina sections were whole mounted on a platinum harp with their photoreceptors 355 facing down, suspended ~1 mm above the glass bottom of the recording chamber. The retina was kept ~32˚C and continuously superfused with Ames media (Sigma-Aldrich, www.sigmaaldrich.com) equilibrated with 95%O2/5%CO2.

Light-stimulation
Light stimuli were generated in Igor Pro 8 (Wavemetrics, www.wavemetrics.com) PC 360 and displayed with a 415 nM LED collimated and masked by an LCD display (3.5 Inch, 480x320 pixels, refresh rate of 50 Hz) controlled by a custom-written python script running on raspberry pi 3 computer. Display luminosity was gamma corrected with a powermeter (Thorlabs, www.thorlabs.com); the stimulus was set to either 60% or -60% Michelson contrast.  Fluorescence signals were collected in a rapid bidirectional frame scan mode (128x64 pixels; 385 ~50 Hz, Thorimage). The line spacing on the vertical axis was doubled to produce a rectangular imaging window (typically ~82x82 µm size, in some experiments, the window was set to ~164x164 µm; the corresponding pixel sizes were 0.64 µm or 1.28 µm). To reduce shot noise, images were subsampled by averaging 2x2 neighboring pixels and filtered by a 20 Hz low pass filter offline. Horizontal and vertical image drifts were corrected online using a reference z-stack 390 acquired before time-series recordings.
For pharmacological manipulations, we used SR95531 (50 µM, Abcam, www.abcam.com) to block GABAA receptors, TPMPA (50 µM, Tocris, www.tocris.com) to block GABAC receptors and strychnine (1 µM), Abcam) to block glycine receptors. All drugs were mixed with the bath Ames medium. We computed the horizontal RF position from responses to motion over the entire display. We first determined the timing of 50% rise-time from trials with leftward and rightward motion. ROIs with their RF center in the middle of the display should respond to both stimuli at the same time following stimulus presentation. In an ROI where the center of the RF is located 410 to the left/right of the display center, a rightward moving stimulus elicits a response that comes earlier/later compared to a trial with a leftward moving stimulus. RF position was computed as half the time difference between the diametrically opposed trials, multiplied by stimulus speed.
Trial responses were considered to be to full-field stimulation if the RF center was at least 100 µm away from the nearest visual edge formed either by masks or the boundaries of the display. 415 Similarly, responses were considered to be near an edge if at least one of the visual edges was closer than 50 µm to the RF center.
To detect similarly shaped groups between different experiments, we conducted a secondary hierarchical clustering. Our initial clustering incorporated responses from trials with moving stimuli and responses near visual edges. Motion responses shift in time as the stimulus 420 progresses over the retina, making comparisons between ROIs difficult. Edge effects may also affect the shape of the responses. For these reasons, as an input to the similarity matrix, we performed a pairwise comparison between 1-second long responses to full-field static stimulation only, for positive contrast stimuli presentation for ON groups and negative stimuli for the OFF groups. As before, the optimal cluster number was determined with the McClain-  Bonferroni correction was used for multiple comparisons. Whenever ratios between parameters were compared, statistics were computed on a logarithmic transformation of the data.

Modeling
All simulations were conducted in Igor Pro 8. The full RF was computed according to the following equation: Where FactorSurround indicated the intensity of the surround activation and varied between 0 (no surround) to 0.5.

455
Simulated neurons were distributed on a 1000 x 1000 µm square grid stimulated either by moving / stationary bars with similar parameters (speed, contrast, size) as in the experiments or by natural images.

Natural movies
The natural movies were composed of background/mask chosen from individual frames 460 of the 'catcam' database 22,52 and stimuli depicting birds of prey. The images were cropped to 100 x 100 pixels and presented as an input to the simulated network. The intensity of the background/mask was scaled to be at the mean pixel level (i.e., 128 pixel luminance value) with an SD of 30. The mean intensity of the stimuli was set to be 2 SD higher than the background mean. In some simulations, the stimulus was not presented. Instead, the background translated

Detailed retinal simulation
The simulated retina consisted of a one-dimensional array (length=700 µm) of photoreceptors, horizontal cells, and BCs, spaced 10 µm apart. Stimuli were provided by a 475 bright bar that was either flashed for 2 seconds or moved over the retina (speed=0.5 mm/s).
Visual edges were created by masking visual presentation near the borders of the array. The simulation time step was 1 ms.
Photoreceptor activation was modeled as a difference between two activation functions (PhA, PhB) with instantaneous rise time and decay times of 60 and 400 ms, respectively. Where the Photoreceptor-horizontal cell gain was set to 1 unless specified otherwise; n=150 is the number of photoreceptors, di,jPhj represents the dimensionality corrected signal from photoreceptor j on horizontal cell i and the last term used to correct responses by RF size.

495
The total activation of the horizontal cells at a time step t was given by the following equation: In which τHC is the horizontal cell activation time constant = 120 ms.
Each photoreceptor combined horizontal cell signals (normalized by the same distance function) with visual illumination as follows: Where the photoreceptor-horizontal cell gain was set to 1, VSi,t represents the value of the visual stimulus over photoreceptor i at time t and HCj,t is the feedback from horizontal cell j.
Similar to horizontal cells, BCs sampled photoreceptor input by dimensionality-corrected RF (size=50 µm unless specified otherwise). The steady-state input-output transformation at 505 the photoreceptor-BC synapse was given by the following relationship: Where di,j was the distance function computed as for horizontal cells, Vslope and V½ defined the slope and the 50% point of the Ph-BC transformation function, and the last term provided a subtraction of the baseline photoreceptive signal.   shown in (a). c Top, the peak response amplitude to stimulus motion from a population of simulated neurons. Activation is maximal near stimulus emergence. Bottom, the mean (±SD) change in RF activation vs. distance from the mask (n=1000 permutations of the background, the horizontal scale is preserved for both plots). Dashed trace, responses in the absence of the mask. d-e As in (b-c), with the surround component removed from the RF description. f The 565 mean(±SD) mutual information computed from the differences in responses of individual neurons located near (<100 µm) the location of stimulus emergence to simulations in the presence or the absence of the mask. ***p<0.001 between the two RF architectures (t-test).