Neuronal activity was recorded from two anesthetized rhesus macaques (M1 and M2, Macaca mulatta) using high-density, multi-contact Neuropixels probes (version 3A, IMEC Inc., Belgium) (Figure 1A) (Methods). These animals had no prior experience with awake experiments or exposure to the stimuli used in this study. Neuropixels probes were inserted into the lateral operculum of V1 with the aid of a surgical microscope at angles nearly perpendicular to the cortical surface. Prior to insertion, the probes were coated with a DiI derivative for subsequent histological visualization of the tracks (Methods) (Supplementary Figure S1). Four Neuropixels probe penetrations were made either in the opercular surface (M1, pen 1-2, ∼4-6° eccentricities) or within the underlying calcarine sulcus (M2, pen 3-4, ∼6-10° eccentricities) using the programmable channel selection feature of the Neuropixels probe. The boundaries of laminar compartments were estimated by combining current-source density (CSD) measurements and histological data (Figure 1B; Supplementary Figure S1) (Methods) (Zhu et al., 2024; Carr et al., 2025). Each recorded neuron was assigned to one of four laminar compartments, specifically 5/6, 4C, 4A/B, 2/3 (mean thickness: 489, 281, 311, 650 µm, respectively, consistent with previous anatomical (O’Kusky and Colonnier, 1982; Lund, 1973) and CSD (Self et al., 2013; Chen et al., 2017) studies). We measured the responses from a total of 677 visually driven neurons (N = 159, 181, 210, and 127 per recording) to classical receptive field (CRF) stimulation, and 621 neurons (N = 142, 179, 152, and 148 per recording) to nonclassical receptive field (nCRF) stimulation (Methods). As a result of the nearly perpendicular penetrations, the visual CRFs of V1 neurons were largely overlapping across the cortical depth (Figure 1C).

During each recording session, neurons were first tested with CRF stimuli consisting of drifting Gabor gratings with 1.5 degrees of visual angle (dva) in diameter, presented at the joint CRF location of simultaneously recorded neurons (Methods). In each of the four recordings, neurons across cortical depth exhibited very similar preferences to grating orientations (Figure 1D). The modal pairwise difference in preferred orientation between neurons across layers was near zero (pen1 = 5°; pen2 = 5°; pen3 = 1°; pen4 = 1°). This result confirmed that our recordings were nearly perpendicular and included neurons predominantly from single orientation columns. Gratings were sized to largely restrict the stimuli within the CRFs of recorded neurons to maximize the extent to which V1 responses were driven in large proportion by feedforward circuitry. In V1, geniculocortical inputs initially drive neurons within layers 4Cα and 4Cβ (Hubel and Wiesel, 1972; Hendrickson et al., 1978; Blasdel and Lund, 1983), which then propagate to supragranular and infragranular layers (Callaway, 1998; Callaway, 2004).

Within each of the recordings, we also examined the responses of V1 neurons to nCRF stimuli. It is known that neurons in primate visual cortex integrate visual information from far beyond their CRFs (Albright and Stoner, 2002; Allman et al., 1985; Fitzpatrick, 2000; Roelfsema, 2006; Angelucci et al., 2017). One example of contextual modulation is the selectivity of neurons to border ownership (B_own). Specifically, neurons in early visual areas respond differently to identical edges (borders) of objects when those objects lie at different locations outside the CRFs (Zhou et al., 2000; Franken and Reynolds, 2021; Hesse and Tsao, 2016; Hesse and Tsao, 2023) (Figure 2). As with other forms of nCRF modulation, B_own is thought to emerge either by feedback from higher areas (Craft et al., 2007; Jehee et al., 2007; Layton et al., 2012; Eguchi and Stringer, 2016; Wagatsuma et al., 2021; Mehrani and Tsotsos, 2021) or from horizontal connections within the same area (Zhaoping, 2005; Kogo et al., 2010). In V1, feedback and horizontal inputs arrive principally within supragranular and infragranular layers, i.e., they avoid layer 4 (Rockland and Pandya, 1979; Rockland and Lund, 1983; Rockland and Virga, 1989; Anderson and Martin, 2009; Markov et al., 2014; Federer et al., 2021; Gilbert and Wiesel, 1983; Shmuel et al., 2005; Siu et al., 2021). Thus, V1 activity generated by nCRF stimulation is known to involve different circuits compared to the activity generated by CRF stimulation.

In this study, border ownership test stimuli consisted of uniformly white or black objects (squares, 8 x 8 dva) on a black or white background, respectively, as in previous studies (Zhou et al., 2000) (Methods). One border of the object fell within the CRFs of the recorded neurons. Within the CRF, the oriented border could either be a dark-light edge or light-dark edge, resulting in two different local contrast polarity (LC) conditions (Figure 2A). Outside the CRF, the stimulus configuration differed for the same LC condition such that the oriented border could belong to opposite sides of an object relative to the CRF. Crucially, across these two object side conditions, the stimulus falling within the CRF was identical (Figure 2B) and was identical within an area that extended well beyond the CRF (8 x 16, dva). In this stimulus configuration, differences in evoked responses between the two object side conditions signify selectivity to border ownership (Zhou et al., 2000). The statistical significance was determined by a two-factor ANOVA on the mean spike counts for each condition, with object side (Side 1 and Side 2) and local contrast (LC1 and LC2) as the factors, along with their interaction (Methods).

We observed neurons with B_own in our recordings from macaque V1. Figures 2C and 2D shows examples of four neurons selective to B_own. Each of these neurons exhibited a preference for LC, with two neurons preferring one contrast polarity and two preferring the other. In addition, each neuron responded differently depending on which side of the object appeared relative to the CRF, thus exhibiting B_own. Three of the neurons responded more vigorously to borders belonging to the lower-left side of object (Side 1) (p < 10^-14, 10^-10, 10^-3 for neurons 1, 2, and 4, respectively), while one neuron responded more vigorously to the border belonging to the upper-right side of the object (Side 2) (p = 0.014, neuron 3). Furthermore, one of the neurons showed a consistent preference for Side 1 across both LCs (neuron 4).

To quantify the B_own, as in previous studies (Zhou et al., 2000), we calculated a B_own modulation index, defined as the difference in neuronal responses to borders belonging to opposite sides of objects, normalized by the maximum response for each neuron (Methods). Responses to the same side of object, but under opposite LC conditions, were averaged together to control for global luminance differences in the overall display. Across the 621 neurons recorded in four sessions, we found that 133 neurons (21.4% overall; pen1 = 28.2%; pen2 = 27.9%; pen3 = 18.4%; pen4 = 8.4%) responded differently to identical borders yet belonging to opposite sides of the object, a proportion consistent with previous studies in awake monkeys (Zhou et al., 2000). Neurons with B_own were found within all cortical laminar compartments, including input layer 4C (5/6: 28/144 = 19.4%; 4C: 48/212 = 22.6%; 4A/B: 19/114 = 16.7%; 2/3: 28/99 = 28.3%) (Table S1), and the proportion was independent of laminar compartment (χ²(3) = 4.78, p = 0.19, chi-quare test of homogeneity). Of all V1 neurons, 96 (15.5%) exhibited B_own depending on the LC (5/6: 11.1%; 4C: 17.9%; 4A/B: 14.0%; 2/3: 20.2%), while 37 (6.0%) showed significant effects independent of local contrast (5/6: 8.3%; 4C: 4.7%; 4A/B: 2.6%; 2/3: 8.1%). Thus, the majority of V1 neurons with B_own exhibited selectivity for only one local contrast, and neurons with B_own invariant to local contrast were rare, also consistent with previous observations (Zhou et al., 2000).

We next considered whether neurons with B_own comprised a distinct group, or instead reflected a general tendency of neurons to prefer one side of the object, with statistically significant neurons falling at the tails of the distribution. We leveraged the large number of simultaneously recorded single neurons to measure the distribution of B_own modulation across laminar compartments in each session. We found that in each of the four recordings, the distributions of B_own modulation generally appeared unimodal, with neurons exhibiting B_own falling at the distribution tails (Figure 2E). Surprisingly, we also found that within individual recordings, neurons across the cortical depth tended to share similar preferences for the side of the object. That is, for each recording, the B_own modulation index differed significantly from 0 (Median modulation index: pen1 = 0.14, p < 10^-12; pen2 = 0.08, p = 0.04; pen3 = −0.18, p < 10^-15; pen4 = 0.10, p < 10^-5, sign test against zero-median). For example, the first recording (pen1) was performed in a column in which most neurons preferred an orientation of 135°. In addition, within that column, neurons tended to exhibit higher responses to a border when it belonged to the lower-left side of an object (side 1), as opposed to the upper-right side (side 2). Thus, V1 exhibited a continuum of B_own, and neurons within single columns tended to prefer the same side of objects located outside of the CRF.

The large number of simultaneously recorded neurons from each session also allowed us to measure interlaminar information flow among single neurons under different visual stimulation conditions. Temporally precise cross-correlations in spike trains offer a unique means of assessing functional interactions among neurons in neural circuits (Perkel et al., 1967). Cross-correlations can be interpreted as suggesting one of myriad putative circuit arrangements among neuronal ensembles (Moore et al., 1970; Aertsen and Gerstein, 1985; Melssen and Epping, 1987; Ostojic et al., 2009). The identification of such interactions has played an important role in elucidating neural circuits in the mammalian visual system (Ts’o et al., 1986; Schwarz and Bolz, 1991; Reid and Alonso, 1995; Alonso et al., 1996; Alonso and Martinez, 1998; Usrey et al., 1998; Nelson et al., 1992; Briggs et al., 2013; Roe and Ts’o, 1999; Senzai et al., 2019; Denman and Contreras, 2014; Siegle et al., 2021; Jia et al., 2022; Trepka et al., 2022; Smith and Kohn, 2008; Sibille et al., 2022). In particular, it has been useful in specifying the flow of signals through local neural circuits (Trepka et al., 2022) and distributed networks (Siegle et al., 2021; Jia et al., 2022; Sibille et al., 2022). Notably, past studies have demonstrated that neuronal cross-correlations are dynamic and can depend on stimulus context and behavioral variables (Hembrook-Short et al., 2019; Martin and von der Heydt, 2015; Hung et al., 2007; Hirabayashi and Miyashita, 2005; Hirabayashi et al., 2010; Steinmetz et al., 2000).

Given that V1 activity generated by nCRF stimulation is thought to involve different circuitry than activity generated by CRF stimulation (Angelucci et al., 2017), we next asked if the pattern of cross-correlations measured under classical and nonclassical stimulus conditions might differ in a manner consistent with models of their underlying circuitry. Specifically, modulation of visual responses during nCRF stimulation, such as border ownership stimuli, is thought to emerge from either feedback from higher areas or horizontal connections within the same areas (Bair et al., 2003; Bullier et al., 2001; Hupe et al., 1998; Chen et al., 2017; Gieselmann and Thiele, 2022; Keller et al., 2020; Klink et al., 2017; Nassi et al., 2013; Nurminen et al., 2018; Angelucci et al., 2017; Angelucci and Bressloff, 2006; Ichida et al., 2007; Schwabe et al., 2006; Pak et al., 2020; Vangeneugden et al., 2019; Zhang et al., 2014; Adesnik et al., 2012; Chisum et al., 2003; Stettler et al., 2002; Liang et al., 2017; Craft et al., 2007; Jehee et al., 2007; Chen et al., 2014; Self et al., 2013; van Kerkoerle et al., 2014), both arriving principally in superficial (layers 1-3) and deep (layers 5-6) layers (Rockland and Lund, 1983; Rockland and Pandya, 1979; Rockland and Virga, 1989; Markov et al., 2014; Federer et al., 2021; Anderson and Martin, 2009; Gilbert and Wiesel, 1983; Shmuel et al., 2005; Siu et al., 2021). In contrast, CRF responses are driven predominantly by feedforward circuitry and geniculocortical inputs to layers 4C (Hubel and Wiesel, 1972; Hendrickson et al., 1978; Blasdel and Lund, 1983; Callaway, 1998; Callaway, 2004). This suggests that the pattern of cross-correlations among neurons across V1 layers may differ under CRF and nCRF stimulus conditions.

Across recording sessions, we calculated cross-correlograms (CCGs) from all possible pairwise combinations of neuronal spike trains during both CRF (Gabor gratings) and nCRF (border ownership) stimulation conditions. Figure 3A shows an example recording session (M1, pen1) in which 159 visually responsive neurons were recorded during CRF stimulation, and CCGs were computed for 7,956 pairwise combinations of neurons (Methods). CCGs were determined to be significant if the jitter-corrected CCG peak occurred within 10ms of zero-time lag, and exceeded 7 standard deviations (SD) above the mean of the noise distribution (Siegle et al., 2021; Trepka et al., 2022) (Methods). In this recording, 13.8% of the total CCGs computed were found to be significant. Eight significant pairs are highlighted in Figure 3A, with their corresponding CCGs shown in Figure 3B. The peak lag (PL), defined as the relative time delay in occurrence of peak correlation between two spike trains, was restricted to within 10ms. This metric approximates the synchrony and/or the direction of information flow between neuronal pairs (Aertsen and Gerstein, 1985; Moore et al., 1970; Melssen and Epping, 1987; Ostojic et al., 2009). These examples illustrate the tendency of CCGs to corroborate the circuitry within V1 columns under CRF conditions in that 1) PLs generally increase with larger distances between pairs, and 2) layer 4 neurons tend to lead superficial neurons (Trepka et al., 2022). We exploited this latter point to test the extent to which the sequence of information flow across layers differs between CRF and nCRF conditions.

From the 677 and 621 neurons recorded across the 4 recording sessions during CRF and nCRF stimulation, there were 41,759 and 28,430 pairwise combinations included, respectively (Methods). Of those, we found that 6,937 (16.6%) and 4,826 (17.0%) of the CCGs were significant (CRF: pen1: 1099/7956 = 13.8%, pen2: 2402/11431 = 21.0%, pen3: 2174/17780 = 12.2%, pen4: 1262/4592 = 27.5%; nCRF: pen1: 447/2473 = 18.1%, pen2: 2750/11823 = 23.3%, pen3: 839/6998 = 12.0%, pen4: 790/7136 = 11.1%). Notably, the proportion of significant connections were nearly identical under the two stimulus conditions, and also were within the range observed in previous studies of macaque V1 (Chu et al., 2014; Hembrook-Short et al., 2019; Smith and Kohn, 2008; Kohn and Smith, 2005).

As noted above, modulation of visual responses during nCRF stimulation, such as border ownership stimuli, is thought to be generated principally by feedback/horizontal connections arriving within superficial and deep layers, which contrasts with feedforward input driving CRF stimulation (Figure 3C). Thus, we asked whether cross-correlations between pairs of neurons situated across laminar compartments depended on the type of visual stimulation. We focused our analysis on CCGs computed from pairs of neurons in which one neuron was located within laminar compartments receiving feedback/horizontal inputs (FH_i), namely layers 2/3 and 5/6, and the other was located within compartments relatively devoid of those inputs, namely 4C and 4A/B. This analysis compared the relative spike timing between neuronal pairs during CRF and nCRF stimulation. If indeed FH_i contribute more to activity during the latter than the former, then we should expect this to be reflected in lead-lag relationships of the CCGs.

A comparison of V1 CCGs from key interlaminar neuronal pairs during CRF and nCRF stimulation is shown for an example recording session (M1, pen1) in Figure 3D. During CRF stimulation, the timing of spikes within compartments 4C and 4A/B generally preceded, or was simultaneous with, spikes within FH_i compartments (CRF: 4C-2/3_PL = 2ms, p < 10^-7; 4A/B-2/3_PL = 2ms, p < 10^-28; 4C-5/6_PL = 0ms, p = 0.52; 4A/B-5/6_PL = 0ms, p = 0.64; Sign test against zero-median). In contrast, during nCRF stimulation, relative spike timing was shifted in favor of neurons in FH_i compartments. Specifically, during nCRF stimulation, the timing of spikes within compartments 4C and 4A/B lagged or was simultaneous with, spikes within FH_i compartments (nCRF: 4C-2/3_PL = 0ms, p = 0.80; 4A/B-2/3_PL = 0ms, p = 0.70; 4C-5/6_PL = −2.5ms, p = 0.11; 4A/B-5/6_PL = 0ms, p = 0.56; Sign test against zero median). To further quantify the apparent shift in lead-lag relationships, we calculated an asymmetry index for each CCG. The asymmetry index is computed as the difference in the CCG integral on either side of zero-time lag, and summarizes the lead-lag relationship independent of the CCG peak (Jia et al., 2022) (Methods). Consistent with the analyses of PL, we found that CCG asymmetries differed significantly between CRF and nCRF stimulation for each of the interlaminar comparisons. In each case, the distribution of asymmetries during nCRF stimulation shifted to favor neurons in FH_i laminar compartments, when compared to CRF stimulation (4C-2/3: CRF_Asym = 0.036, nCRF_Asym = −0.004, Δ = −0.04, p < 10^-6; 4A/B-2/3: CRF_Asym = 0.046, nCRF_Asym = −0.005, Δ = −0.051, p < 10^-11; 4C-5/6: CRF_Asym = −0.010, nCRF_Asym= −0.044, Δ = −0.034, p = 0.0015; 4A/B-5/6: CRF_Asym = 0.012, nCRF_Asym = −0.028, Δ = −0.04, p < 10^-5; Wilcoxon rank-sum test). Thus, CCG data from this single recording session indicated that relative spike timing was earlier within FH_i laminar compartments during nCRF stimulation than during CRF stimulation.

Across all recordings, we identified a total of 2,573 and 2,118 significant neuronal pairs from the four interlaminar combinations during CRF (pen1 = 274; pen2 = 1187; pen3 = 560; pen4 = 302) or nCRF stimulation (pen1 = 210; pen2 = 1226; pen3 = 382; pen4 = 300), respectively, and compared their CCG asymmetries. Similar to the pattern observed in the example recording, we observed a consistent shift in lead-lag relationships for each of the interlaminar comparisons (Figure 4). Specifically, during nCRF stimulation, the distribution of asymmetries during nCRF stimulation shifted to favor neurons in FH_i laminar compartments, when compared to CRF stimulation (4C-2/3: CRF_Asym = 0.028, nCRF_Asym = 0.012, Δ = −0.016, p < 10^-9; 4A/B-2/3: CRF_Asym = 0.038, nCRF_Asym = 0.022, Δ = −0.016, p < 10^-9; 4C-5/6: CRF_Asym = −0.033, nCRF_Asym = −0.047, Δ = −0.014, p < 10^-5; 4A/B-5/6: CRF_Asym = −0.018, nCRF_Asym = −0.05, Δ = −0.032, p < 10^-20; Wilcoxon rank-sum test). Thus, relative spike timing was earlier within FH_i laminar compartments during nCRF stimulation than during CRF stimulation in the combined dataset, suggesting that the proportion of information flow from FH_i layers to input layers increased during nCRF stimulation. This observation was consistent across individual recording sessions (Table S2). Notably, we observed a similar pattern when instead of asymmetry, we compared CCG peak lags (Supplementary Figure S2A), and when we compared the total number of both significant and nonsignificant pairs (CRF = 15,920; nCRF = 11,263) (Supplementary Figure S2B), and when we compared CCGs with a more stringent significance criterion (10 SDs) in the combined dataset (Supplementary Figure S2C-D).

Lastly, we asked whether interlaminar information flow was related specifically to border ownership modulation, independent of the comparisons between CRF and nCRF stimulation. Since CRF and nCRF stimuli differed in ways potentially unrelated to nonclassical modulation (e.g., variations within the CRF), we sought to assess the extent to which information flow from FH_i to input layers predicted the magnitude of nCRF modulation. To this end, we computed two modulation indices from the same B_own test stimuli (Figure 2A), specifically the B_own modulation index and the local contrast (LC) index. In contrast to the B_own modulation index, which reflects nCRF modulation, the LC index measures selectivity to luminance contrast polarity (dark-light vs. light-dark) within the CRF. We hypothesized that if the relative timing of spiking activity within FH_i laminar compartments is indicative of nCRF modulation, we should expect earlier relative spike timing in FH_i compartments, i.e., lower CCG asymmetries with respect to other compartments, to be associated with greater B_own. In contrast, we should not expect that relationship with the LC index.

Consistent with our hypothesis, we observed a negative relationship between the B_own modulation index and CCG asymmetry across the population and within individual sessions (linear regression: population slope (β) = −0.32, p < 10^-27; individual sessions: pen1 = −0.52, p < 10^-5; pen2 = −0.06, p = 0.14; pen3 = −0.25, p < 10^-4; pen4 = −0.20, p = 0.008) (Figure 5A). This result suggests that the magnitude of B_own modulation for neuronal pairs depended on the proportion of information flow from FH_i compartment to input layers. In contrast, the opposite was true for the LC index. That is, LC index was positively correlated with CCG asymmetry across the population (linear regression: population slope (β) = 0.46, p<10^-28) (Figure 5B). However, this relationship was not consistent across sessions (individual sessions: pen1 = −0.30, p = 0.09; pen2 = 0.33, p < 10^-5; pen3 = 0.23, p = 0.006; pen4 = 0.14, p = 0.04). To explore both results further, we built generalized linear models (GLMs) to assess the dependence of B_own or LC modulation on CCG asymmetry to incorporate variations across interlaminar combinations and recording sessions. For LC modulation, this analysis indicated that the magnitude of LC modulation was not reliably predicted by interlaminar information flow (GLM: coefficient = −0.04, p = 0.30; R² = 0.27) (Table S3). In contrast, for B_own modulation, this analysis confirmed that the magnitude of modulation was reliably predicted by interlaminar information flow, independent of variations across interlaminar combinations and recording sessions (GLM: coefficient = −0.16 p = 1.1*10^-6; R² = 0.11). These results demonstrate that interlaminar information flow was related specifically to the magnitude of border ownership modulation.

Anesthetized recordings were conducted in two adult male rhesus macaques (Macaca mulatta, M1, 13 kg; M2, 8 kg). All experimental procedures were in accordance with the National Institutes of Health Guide for the Care and Use of Laboratory Animals, the Society for Neuroscience Guidelines and Policies, and with approved Institutional Animal Care and Use Committee (IACUC) protocol (#APLAC-9900) of Stanford University.

Prior to each recording session, treatment with dexamethasone phosphate (2 mg per 24 h) was instituted 24 h to reduce cerebral edema. After administration of ketamine HCl (10 mg per kilogram body weight, intramuscularly), monkeys were ventilated with 1-2% isoflurane in a 1:1 mixture of N₂O and O₂ to maintain general anesthesia. Electrocardiogram, respiratory rate, body temperature, blood oxygenation, end-tidal CO₂, urine output and inspired/expired concentrations of anesthetic gasses were monitored continuously. Normal saline was given intravenously at a variable rate to maintain adequate urine output. After a cycloplegic agent (atropine sulfate, 1%) was administered, the eyes were focused with contact lenses on an LCD monitor. Vecuronium bromide (60 μg/kg/h) was infused to prevent eye movements.

With the anesthetized monkey in the stereotaxic frame, an occipital craniotomy was performed over the opercular surface of V1. The dura was reflected to expose a small (∼3 mm²) patch of cortex. Next, a region relatively devoid of large surface vessels was selected for implantation, and the Neuropixels probe was inserted with the aid of a surgical microscope. Given the width of the probe (70 µm x 20 µm), insertion of it into the cortex sometimes required multiple attempts if it flexed upon contacting the pia. The junction of the probe tip and the pia could be visualized via the (Zeiss) surgical scope, and the relaxation of pia dimpling was used to indicate penetration, after which the probe was lowered at least 3-4 mm. Prior to probe insertion, it was dipped in a solution of the DiI derivative FM1-43FX (Molecular Probes, Inc) for subsequent histological visualization of the electrode track.

Given the length of the probe (1 cm), and the complete distribution of electrode contacts throughout its length, recordings could be made either in the opercular surface cortex (M1, pen 1-2) or within the underlying calcarine sulcus (M2, pen 3-4), by selecting a subset of contiguous active contacts (n = 384) from the total number (n=986). Recordings were made at 1 to 2 sites in one hemisphere of each monkey. At the end of the experiment, monkeys were euthanized with pentobarbital (150 mg/kg) and perfused with normal saline followed by 1 liter of 1% (wt/vol) paraformaldehyde in 0.1 M phosphate buffer, pH 7.4.

Visual stimuli were presented at manually mapped receptive field (RF) locations for each recording session on an LCD monitor (Model NEC-4010, dimensions: 88.5 cm (H) x 49.7 cm (V), resolution: 1360 x 768 pixels, frame rate: 60 Hz) positioned 114 cm from the monkeys. RF eccentricities were ∼4-6° (M1) and ∼6-10° (M2). Visual stimuli were generated using customized MATLAB scripts with the Psychophysics Toolbox extensions (version PTB-3) (Kleiner et al., 2007). A photodiode was used to measure stimulus timing.

Classical receptive field (CRF) stimuli consisted of drifting Gabor gratings (2 deg./sec., 100% Michelson contrast) with a diameter of 1.5 degrees of visual angle (dva), positioned within the joint receptive fields (RFs) of recorded neurons. This size was selected to largely constrain the stimuli within the CRFs of recorded neurons, typically ∼0.5-1 dva, as determined manually online. Gratings drifted in 36 different directions, from 0 to 360° in 10° steps, in a pseudorandom order. Each stimulus condition was presented for 1 s and repeated 5 or 10 times. A blank screen with equal luminance to the Gabor patch was presented for 0.25 s during the stimulus interval. Stimuli were presented either monocularly (pen 1, 3) or both monocularly and binocularly (pen 2, 4). Four spatial frequencies (0.5, 1, 2, 4 cycles/deg.) were tested. The optimal eye and spatial frequency conditions were determined offline for further analysis.

Nonclassical receptive field (nCRF) stimuli used in this study, i.e., Border ownership test stimuli consisted of uniformly white or black objects (square, 8 x 8 dva) presented on a black or white background, respectively, like previous studies (Zhou et al., 2000). One border of the object fell within the CRFs of the recorded neurons at an orientation that varied in 45° steps. Objects were either 4 x 4 dva or 8 x 8 dva in size, and only conditions with 8 x 8 dva were selected for quantifying border ownership selectivity to tolerate the variation in the CRF locations. Stimuli were presented monocularly to the preferred eye, to avoid the surround stimulating the CRF of the non-preferred eye. For each of the 4 orientation conditions, there were 4 basic conditions (Figure 2A). Within the CRF, the oriented border could either be a dark-light edge or light-dark edge, resulting in two different local contrast (LC) polarity conditions, namely, LC1 and LC2. For the same LC condition, the stimulus configuration outside of the CRF differed such that the oriented border could belong opposite sides of objects, resulting in two different sides of object conditions, namely, side 1 and side 2. Each stimulus condition was presented for 1 s and repeated 10 or 20 times. A blank screen with equal luminance to the mean of object and background luminance was presented for 0.5s during the stimulus interval.

In some sessions, the receptive fields of V1 neurons were tested systemically for offline use. CRFs were mapped from a 10 x 10 array (4 x 4 dva) using probe stimuli. The array was centered at the manually determined RFs location. The stimuli consisted of circular sine wave gratings of 0.4 dva in diameter, and 100% Michelson contrast. Gratings were presented at the optimal orientation determined manually online for each recording session and drifted at speed of 2 deg./sec. Each stimulus condition was presented for 0.2s, interleaved with 0.3s of blank screen with equal luminance, and repeated 15 times.

Mean spike counts from 0.05 to 0.25s after each stimulus onset were calculated, resulting in a 10 x 10 matrix containing mean spike rates to each grid location. This 2D matrix was then interpolated to a 40 x 40 matrix (MATLAB function ‘interp2’) and then smoothed with a 2D Gaussian filter (MATLAB function ‘imgaussfilt’, σ=2). The outline of the CRF was defined as the contour isoline at 80% of the maximum response using MATLAB function ‘contour3’. Neurons with only 1 contour region at the threshold were selected.

Raw spike-band data were sampled and recorded at 30kHz. They were then median-subtracted and high-pass filtered at 300 Hz during the pre-processing stage. Spike sorting was carried out with Kilosort 3 (Pachitariu et al., 2024) to find spike times and assign each spike to different templates (neurons). Default parameters in Kilosort 3 were used for spike sorting, specifically, Ops.th = [9,9]; Ops.lam = 20; Ops.AUCsplit = 0.8; Ops.ThPre = 8; Ops.sig = 20; Ops.nblocks = 5; Ops.spkTh = −6. The raw sorted data were then manually curated in Phy (https://github.com/cortex-lab/phy) to remove templates with very few spikes or with atypical waveforms and to perform minimal templates merging and splitting. Double-counted spikes from the algorithm fitting the residuals to a new template were identified using established criteria (Siegle et al., 2021), by counting spikes from two templates separated by less than 50μm (∼5 channels) and with spike activity occurring within 5 samples (0.167ms). Double-counted spikes from the template smaller in amplitude were removed from further analysis. Visual responsiveness of each unit was assessed under CRF, nCRF, or RF mapping stimulation. For each type of stimulation, a paired-sample t test was performed to determine whether the mean spike counts after each stimulus presentation exceeded the preceding blank period at a significance level of 0.01. Only neurons showing significant visual responsiveness were included for further analysis.

The laminar locations of our recorded neurons were estimated based on a combination of functional analysis and histology results (Zhu et al., 2024; Carr et al., 2025). For each recording, we first performed the current source density (CSD) analysis on the stimulus-triggered average of local field potentials (LFP). LFP were low-pass filtered at 200 Hz and recorded at 2500 Hz. LFP signals recorded from every 4 nearby channels were averaged and realigned to the onset of visual stimulus. CSD was estimated as the second-order derivatives of signals along the probe axis using the common five-point formula (Nicholson and Freeman, 1975). The result was then smoothed across space (σ = 120 μm) to reduce the artifacts caused by varied electrode impedance. We located the lower boundary of the major sink (the reversal point of sink and source) as the border between layer 4C and layer 5/6. We also considered anatomical data in order to localize recorded neurons within gray matter, allowing for minor adjustments (±1 group channel) in layer boundary placement. Subsequent layer boundaries were determined by offsetting the cortical thickness derived from histological slices (Supplementary Fig S1).

To assess the orientation tuning for each neuron, the spike rates during 0.1 to 1s after each stimulus onset were calculated and averaged for each orientation condition (N= 18). The orientation tuning responses were first smoothed with a Hanning window (half-width at half-height of 20^°), and then fitted with a von Mises function (Swindale, 1998) The location of the peak of the fitted curve was determined as the preferred orientation. Only neurons well fit by the function (R²>0.7) were included for determining pairwise differences in preferences for all combinations in each recording session.

As in previous studies (Zhou et al., 2000), we calculated a B_own modulation index to quantify the selectivity of each neuron for B_own. The B_own modulation index is defined as Here, R_{LC, Side} represents the average spike rate from 0.05 to 0.5 s after stimulus onset for one of four basic conditions, which are derived from the combination of two different Local Contrast conditions (LC1 and LC2) and two Side conditions (Side 1 and Side 2). Baseline activity was not subtracted. Since the border within the CRF can vary in 45° steps from 0° to 180°, we only selected the optimal orientation for each recording population to calculate the B_own index. R_max was defined as the maximum response for each neuron across the four basic conditions and four orientations. We used R_max instead of summing responses across the four basic conditions to account for the slight variations in orientation preferences among neurons within each recording session. The statistical significance of B_own was evaluated using a two-factor ANOVA (MATLAB function ‘anovan’) on the average spike rates for each condition. The two factors were object Side (with two levels: Side 1 and Side 2) and Local Contrast (with two levels: LC 1 and LC2), including their interactions. A significance level of 0.05 was used. Neurons were considered to exhibit significant B_own modulation if they showed a significant main effect for the object Side factor. Whether they were invariant to local contrast was determined by the main effect of the Local Contrast factor.

To measure correlated firing, we computed the cross-correlations between spike trains of all pairs of simultaneously recorded neurons. For each recording session, we calculated cross-correlograms (CCGs) from pairwise combinations of neuronal spike trains during either CRF (Gabor gratings) or nCRF (border ownership) stimulation conditions, respectively. We only included neuronal pairs with geometric mean firing rates more than 0.5 spikes/sec for further analysis. For both conditions, we focused on the spiking activity within the 0.25 −1s window during each stimulus presentation, to mitigate the influence of the transient visual response after stimulus onset. The CCG function for a given pair of neurons was defined as follows: where j denotes the first/reference neuron and k denotes the second/target neuron in the CCG function. τ is the time lag between the two spike trains and from neuron j and k, during trial i. The value of and is 1 if there is a spike at time bin t and is zero otherwise. M is the total number of trials; N is the number of time bins within a trial. θ(τ) = N − | τ|, is a triangular function that corrects for the degree of overlap in the two spike trains at each time lag. λ₉ and λ: are the mean firing rates of neuron j and k computed over the same bins used to compute the CCG at each time lag. The normalization by the geometric mean of spike rates is commonly used by previous studies to account for the influence of firing rates on the CCG peaks (Bair et al., 2001; Kohn and Smith, 2005).

To correct for correlations due to stimulus-locking or slow fluctuations in the population responses (e.g., gamma-band activity), we computed a jitter-corrected CCG by subtracting a jittered CCG from the original CCG: The jittered CCG (CCG_jittered) reflects the expected value of CCG computed from all possible jitters of each spike train within a given jitter window (Harrison and Geman, 2009; Smith and Kohn, 2008). The jittered spike train preserves both the PSTH of the original spike train across trials and the spike counts in the jitter window within each trial. As a result, jitter correction removes the correlation between PSTHs (stimulus-locking) and correlation on timescales longer than the jitter window (slow population correlations). Here, a 25-ms jitter window was chosen based on previous studies (Jia et al., 2013; Siegle et al., 2021). Jitter-corrected CCG (CCG_corrected) was then smoothed with a 5ms kernel [0.05, 0.25, 0.4, 0.25, 0.05] (Kohn and Smith, 2005; Smith and Kohn, 2008) before further analysis.

We classified a CCG as significant if the peak of the jitter-corrected CCG occurred within 10 ms of zero and exceeded 7 standard deviations (SD) above the mean of the noise distribution, similar to previous studies (Siegle et al., 2021). The noise distribution for a CCG was defined as the flanks of the jitter-corrected CCG ({CCG(τ)|100 ≥ | τ| ≥ 50 ms}). All analyses presented in the main text involve only significant, jitter-corrected CCGs. In the supplementary material, we also applied a more stringent criterion, requiring that significant CCG peaks exceed 10 SD above the noise, and a less stringent criterion that includes both significant and non-significant CCGs. These ensure that our results are robust across the choices of threshold (Supplementary Figure S2). To assess the directionality of functional interactions or signal flow between pairs of neurons, we focused on two metrics derived from the CCG functions. One is the peak lag (PL), which describes the time lag of peak correlation between the spike trains of neuron k (target) relative to neuron j (reference). PL is positive if the reference neuron j leads target neuron k in spike timing, negative if j lags k, and zero if j and k fire synchronously. The second metric we used to assess the signal flow is CCG asymmetry, computed by subtracting the sum of CCG values during the [-13, 0] ms time window from the sum of CCG values during the [0, 13] ms. Similar to PL, a positive CCG asymmetry indicates that reference neuron j leads k, while a negative asymmetry indicates that j lags k. The 13ms window was chosen as half of the 25ms jitter window, consistent with previous studies (Jia et al., 2022). Compared to peak lag, CCG asymmetry additionally captures the strength of the functional interactions between pairs of neurons and is less dependent on the exact shape of the CCG peak, which is often more complex in corticocortical functional interactions (Alonso and Martinez, 1998).

To determine whether the magnitude of border ownership or local contrast modulation is predicted by interlaminar signal flow during nCRF stimulation, we first performed linear regression (MATLAB, ‘fitlm’, ordinary least squares) and then built generalized linear models (GLMs) using predictors including CCG asymmetry, interlaminar combinations, and recording sessions. We focused on the four key interlaminar combinations (4C-2/3, 4A/B-2/3, 4C-5/6, 4A/B-5/6), where the latter laminar compartments receive feedback/horizontal inputs (FH_i). Only neuronal pairs with significant CCGs were included as individual samples in the GLM. The resulting GLM equations were: B_own (pair) and LC (pair) were calculated as the geometric mean of the B_own and LC indices for the constituent neurons in each pair. Specifically, , where abs(B_own(j)) and abs(B_own(k)) represent the absolute B_own modulation indices for constituent neurons j and k. A similar calculation was performed for LC (pair). CCG asymmetries were calculated from the nCRF stimulation condition where neurons in the input layers were selected as reference neurons and the ones in FH_i compartments were target neurons in the CCG. Additional predictors, including interlaminar combinations and recording sessions, were included in the models to incorporate variations when pooling the dataset.

B_own for each neuron was assessed with ANOVA. The proportion of B_own neurons within different layers was assessed using chi-squared test of homogeneity. The sign of populational border ownership preferences from individual recording session was tested with sign test. The sign of populational CCG peak lag for a given laminar compartment combination was assessed with sign test. The comparisons of peak lag and CCG asymmetry between CRF and nCRF stimulation condition for each laminar compartment combination were evaluated using Wilcoxon rank-sum tests. The dependence between CCG asymmetry and B_own for pair of neurons was assessed using linear regression and GLM.