Using coherent plaids, we have previously demonstrated signatures of motion integration in PSS of adult ferrets (Figure 1A). Specifically, a group of PSS neurons responds to the pattern direction of the plaids (so called ‘pattern neurons’), rather than the motion of the individual components (as would be the case for so called ‘component neurons’). Here, we first aimed to establish the developmental timeline for PSS pattern responses. To this end, we used tetrodes or multi-channel silicon probes to record PSS responses in anesthetized ferrets at various ages. Previous research in primates suggests that motion integration develops after maturation of direction selectivity (Chino et al., 1997; Kiorpes and Movshon, 2014). In addition, a minimal level of direction selectivity is required for a meaningful analysis of responses to plaids. We therefore restricted our experiments to an age range after V1 direction selectivity development, which in ferrets occurs around P37 (ferrets open their eyes at about P30). Our data confirm that PSS direction selectivity is also mature at this age (see Figure 1—figure supplement 1).

Figure 1 with 1 supplement see all

Download asset Open asset

PSS motion integration was assessed at different ages by probing neurons with plaids and gratings. Since plaids are constructed from two component gratings moving in different directions, each plaid can be described by two parameters (Figure 1B) – the difference in direction between the two components (dOri) and the resulting plaid direction (which bisects the two component directions). Importantly, the perception of a coherently moving plaid is maintained over a large range of dOri values (Adelson and Movshon, 1982). To efficiently sample a large number of dOri values and plaid directions, we presented stimuli in a streaming stimulus paradigm, in which stimuli were shown in rapid succession (stimulus update rate: 3 Hz; Figure 1B). We have previously shown that the results of this approach generally agree with those based on showing single stimuli more slowly, but that the sampling of a larger stimulus space noticeably improves the discriminability of pattern and component responses, in addition to generating results that are independent of a particular dOri choice (see Rust et al., 2006; Smith et al., 2005 for a similar presentation method).

For every neuron probed with the streaming stimulus paradigm, we then determined its capacity to integrate local motion signals (see Lempel and Nielsen, 2019 and Materials and methods for details). Briefly, for every neuron, we generated two predictions from the measured stimulus responses, one for pattern and one for component responses. Each prediction summarized tuning as a two-dimensional (2D) function of both the (integrated) plaid direction and dOri. Since coherent plaid motion is perceived independent of dOri, the pattern prediction assumed that the tuning peak would always occur for plaids moving in the preferred direction, regardless of dOri value (but allowing changes in peak height in a dOri-dependent fashion). Consequently, tuning profiles for pattern predictions had a single ridge along the preferred direction (Figure 1C, top). The component prediction, on the other hand, assumed that the individual component directions determined responses, not the integrated plaid direction. For each dOri, the strongest responses therefore should occur whenever one of the components moved in the preferred direction, resulting in a characteristic ‘V’-shape in the chosen 2D tuning plot format (Figure 1C, bottom). As for the pattern prediction, dOri-dependent changes in maximum rates were included in the component prediction. We then compared the two predictions to the measured tuning profile of each neuron by computing Z-corrected partial correlations, resulting in a pattern correlation (Z_P) and a component correlation (Z_C). We also computed a pattern index for every neuron as Z_P–Z_C. Based on established criteria (Smith et al., 2005), neurons were classified as pattern neurons if Z_P was significantly higher than Z_C (using p<0.1); component neurons met the opposite criterion (see Figure 2C and Materials and methods for precise definitions of category boundaries).

Figure 2 with 1 supplement see all

Download asset Open asset

Figure 2—source data 1

Neuron-by-neuron metrics for Figure 2 and Figure 2—figure supplement 1.

https://cdn.elifesciences.org/articles/59798/elife-59798-fig2-data1-v2.xlsx

Download elife-59798-fig2-data1-v2.xlsx

Across all ages tested (P37–13 months), we were able to observe pattern and component cells in PSS (Figure 1D,E). Nonetheless, the degree of PSS motion integration underwent pronounced changes with age. As a first step in assessing the time course of motion integration development, we computed the median pattern index for every animal. Figure 2A shows these medians as a function of age (Note that for some animals less than three neurons were recorded. We included these animals for completeness but indicate the corresponding data points in Figure 2A.) While this plot confirmed that a range of pattern indices could be observed at every age, it also showed that the lowest median pattern indices generally occurred in animals before P41. After P41, the range of median pattern indices across animals of the same age was generally similar to that of adults. We therefore divided all developmental data into two groups, P37–40 and P41–47. A third group consisted of adult animals (age range P100–448). As expected, the per-animal pattern indices in the P37–40 age groups were significantly lower than those of the other two groups (P37–40 vs P41–47: p=0.002; P37–40 vs Adult: p=0.001. See Table 1 for the full test details and all statistical results. All data selection criteria are outlined in the Materials and methods. Uneven sample sizes were considered in the choice of statistical tests and are discussed in Materials and methods as well).

Next, we assessed the motion integration capacity of these three groups at the population level by comparing responses across all of the neurons collected for each group (rather than the animal-by-animal analysis used for Figure 2A). Figure 2B plots the cumulative distribution of pattern indices for each group. Consistent with the trends shown in Figure 2A, the pattern index increased significantly between P37–40 and P41–47, at which point it became adult-like (P37–40 vs P41–47: p<0.001. See Table 1 for the other p values. Non-significant results are omitted from the results text, but are listed in Table 1). This change in pattern index was driven by opposite changes in both of the underlying partial correlation coefficients (Figure 2—figure supplement 1): Between the two younger age groups, Z_C decreased significantly (p<0.001), while Z_P increased (p<0.001). Changes in pattern and component cell proportions show the same trends (Figure 2C): The proportion of component cells decreased with age (75% at P37–40 vs 41% in adults), while the proportion of pattern cells increased (13% at P37–40 vs 41% in adults).

A possible concern regarding these findings is the short presentation time per stimulus in the streaming stimulus paradigm. Fast stimulus presentation times combined with potentially slower motion integration in younger animals might result in pattern integration remaining undetected. In addition, the rapid succession of stimuli in the streaming paradigm might differentially affect PSS across development. As a control, we therefore recorded responses to plaids and gratings presented individually and for longer durations (1 s) at P37–40, at P41–47, and in adulthood. Plaids were restricted to a dOri value of 135 deg in these experiments (see Plaids, classic stimulus presentation in Materials and methods for full stimulus description). The control experiment confirmed our main findings with the streaming stimulus paradigm (Figure 2D–F): As before, the pattern index increased significantly between P37–40 and P41–47 (p<0.001), at which point it became adult-like. The number of pattern cells followed the same trend. Thus, the initially lower motion integration levels at P37–40 are not simply a consequence of the streaming stimulus paradigm.

Another possible concern regarding the data is that weaker responses in younger animals may mask motion integration. More precisely, higher variability in responses, particularly at younger ages, might make pattern and component correlation indices more similar, thereby lowering the absolute value of the pattern index. Indeed, for plaids with dOri = 90 deg (which generally evoked strong responses; see section on Relative plaid response strength below) the median response rate was 2.46 Hz at P37–40, but 2.97 Hz in the P41–47 group (rank-sum test, p=0.02). To test whether these response rate differences could contribute to the low pattern indices in the youngest age group, we subsampled the P37–40 data set by selecting only the neurons with firing rates higher than the median at P41–47 (i.e., firing rates larger than 2.97 Hz). For the 56 neurons that fell into this category, the median pattern index remained low at −6.03 (versus −5.6 in the full data set). The significant difference in pattern index between P37–40 and P41–47 also remained for the subsampled data set (Welch’s t-test, p<0.001). It therefore seems unlikely that low response rates can explain the lack of motion integration in the youngest age group.

In summary, PSS motion integration – as reflected in the responses to the pattern motion of coherent plaids – developed after direction selectivity and matured about 7 weeks after birth or 1 week after direction selectivity (which also corresponds to about 2 weeks after eye opening).

Visual experience has been shown to play a key role in the development of direction selectivity in ferret V1 (Li et al., 2006; Popović et al., 2018; Van Hooser et al., 2012). Here, we made use of inter-animal variability in development to begin to probe whether it might similarly impact the development of higher-level motion functions in ferrets. In the previous section, we compared PSS responses strictly by age. However, the age at eye opening – and with it the starting point of patterned visual experience – varies across animals (for this data set: range P31-36). This made it possible to test whether differences in the duration of visual experience resulted in measurable differences in motion integration in animals of the same age. Systematic changes in motion integration due to the amount of visual experience should be most detectable in the youngest age group (P37–40), which have the most immature levels of PSS motion integration, and consequently the largest capacity for improvement. We therefore restricted the analysis to animals of this age range. No animal in this group had less than 4 days of experience. A single animal had 7 days of visual experience at the time of recording and was excluded from further analysis as an outlier. Figure 3A plots the median pattern index for each of the remaining animals in the P37–40 age group as a function of their amount of visual experience at the time of recording (same format as for Figure 2A). This figure shows an increase in pattern index with visual experience. In particular, kits with 4 days of visual experience had lower pattern indices than kits with 5 or 6 days of visual experience (V4 vs V5: p=0.004; V4 vs V6: p=0.02). We therefore divided data into two groups, V4 and V5–6, for a population-level analysis (Figure 3B–D). Longer visual experience indeed resulted in a larger proportion of pattern cells (V4: 6%, V5–6: 13%) and a significant increase in pattern index (p=0.01). Analyzed separately, both pattern and component correlations showed significant changes (Z_P: p=0.03; Z_C: p=0.02).

Figure 3

Download asset Open asset

Figure 3—source data 1

Neuron-by-neuron metrics for Figure 3.

https://cdn.elifesciences.org/articles/59798/elife-59798-fig3-data1-v2.xlsx

Download elife-59798-fig3-data1-v2.xlsx

The above analysis relies entirely on natural variation in eye opening across animals. While this offers the advantage of studying the impact of visual experience outside of artificial manipulations, it limits the available data set. For example, certain combinations of gestational age and visual experience duration did not occur in our data set. In addition, some correlation between eye opening and gestational date is unavoidable and resulted in the fact that animals in the V4 group were slightly younger (mean age 37.6 days, range 37–40) than animals in the V5–6 group (mean age 38.6, range 37–40). We therefore tested whether visual experience indeed had an effect in addition to age by performing a two-way ANOVA on the pattern index data from all PSS cells with factors visual experience (V4–V6) and age (P37–40). Main effects for both factors were significant (main effect age: p=0.02; main effect visual experience: p=04), demonstrating that both contributed significantly to the observed variance in pattern index across cells. Due to the sampling limitations imposed by natural variability, additional experiments will be needed to fully disentangle the impacts of age and experience. Yet, our results suggest that visual experience impacts the development of PSS motion integration, in addition to the contribution of age.

Our stimulus set allowed us to probe PSS plaid responses not just as a function of plaid direction, but also as a function of dOri, the intersection angle between the two component gratings. Possible reasons for a response modulation with dOri include differences in the spatial frequency of these plaids, as well as differences in their speed (Zaharia et al., 2019; Priebe et al., 2003). Additionally, the motion integration mechanisms acting in PSS will likely have an impact on plaid response strength by determining how strongly different component signals interact, an assumption that is supported by the observation that the strength of plaid responses is positively correlated with pattern index in both primate MT and adult ferret PSS (Lempel and Nielsen, 2019; Wang and Movshon, 2016).

Here, we found that the dependency of response strength on dOri underwent a series of developmental changes. Figure 4A captures these changes by plotting the population tuning profile as a function of dOri and plaid direction for the different age groups, with the plaid responses of every neuron normalized by the response to gratings. We also quantified changes in relative plaid responses by averaging across all plaid directions (Figure 4B). Note that when analyzed separately for each dOri, the relative plaid response followed a χ² distribution (Figure 4—figure supplement 1). We therefore used a Wilson–Hilferty transformation (Wilson and Hilferty, 1931) to transform relative plaid responses to a normal distribution before any of the subsequent analyses. Figure 4B shows the transformed data; the raw data and effects of transformation are presented for reference in Figure 4—figure supplement 1.

Figure 4 with 1 supplement see all

Download asset Open asset

Figure 4—source data 1

Neuron-by-neuron metrics for Figure 4 and Figure 5—figure supplement 1.

https://cdn.elifesciences.org/articles/59798/elife-59798-fig4-data1-v2.xlsx

Download elife-59798-fig4-data1-v2.xlsx

Comparing the average relative plaid responses across age groups revealed that their dOri dependency systematically changed, especially between the youngest age group and both older groups. Most notably, in the youngest animals, changes in dOri caused less changes in plaid responses than in older animals. This was reflected in significant interactions between dOri and age when comparing the youngest animals against either of the other two groups using a two-way ANOVA (P37–40 vs P41–47, p=0.006; P37–40 vs Adult, p=0.004). As a post hoc test, we compared relative plaid responses between age groups at the dOri values evoking maximal and minimal responses in adults (dOri = 45 and 157 deg, respectively). At dOri = 45 deg, we observed significantly lower plaid responses in the youngest animals than in adults (p=0.005). Responses were not different from the P41–47 group for this dOri value. In contrast, responses at dOri = 157 deg were significantly higher in the youngest group than either of the other groups (P37–40 vs P41–47, p=0.02; P37–40 vs Adult, p=0.01). Thus, the post hoc tests confirm that the youngest animals showed less modulation of plaid response strength with dOri than older animals. It should be pointed out that the reduction in responses to plaids with large dOri values between P37–40 and P41–47 correlated with a large reduction in component responses (Figure 4A). In the youngest age group, the impact of the stronger component responses is visible in the clear ‘V’-shape of the population tuning profile. As could be expected, component responses were less affected by dOri, which resulted in a higher response to plaids with large dOri in the youngest age group when averaging across all plaid directions (as done for Figure 4B).

As mentioned above, the maximum relative plaid response is positively correlated with the pattern index in ferret PSS and primate MT in adult animals (Lempel and Nielsen, 2019; Wang and Movshon, 2016). This relationship also emerged with development (Figure 4C; note that these data also were transformed using the Wilson-Hilferty transformation): While the correlation coefficient between pattern index and maximum relative plaid response was significant even for the P37–40 group, Figure 4C shows the dependency between the two variables was quite weak at this age. The correlation coefficient in the older age groups was significantly higher (P37–40 vs P41–47, p=0.02; P37–40 vs Adult, p=0.003), with a more obvious relationship between the two variables apparent in adults (Figure 4C). In summary, not only did the prevalence of pattern responses to plaids increase between P37–40 and P41–47, the dependency of this response on plaid configuration and pattern index developed as well, showing more comprehensive changes in how the visual system handles stimuli consisting of multiple interacting components.

Processing of stimuli in V1 is likely an important precursor to PSS functions. V1 does not just compute the local direction signals for the plaid components, the response to these components is modulated by interactions between the plaids through mechanisms like cross-orientation inhibition and motion opponency. These V1 mechanisms could further drive modulations in plaid response strength, in addition to the PSS mechanisms discussed above. We therefore sought to identify developmental changes in V1 plaid processing to gain a more complete picture of the development of the motion pathway as a whole. Accordingly, we recorded V1 responses to plaids at different ages. We selected two age groups for these experiments. The first covered the same age range as the youngest PSS animals, P37–40. The second group covered P44–47, chosen to coincide with the end of the age range of the second PSS group so as to maximize our ability to detect age-dependent differences.

Recording V1 responses to plaids revealed a series of developmental changes. First, we observed changes in the strength of plaid-evoked responses (Figure 5A,B). Between the two younger age groups, responses to plaids generally increased independent of dOri (two-way ANOVA, contribution of age p=0.001), before decreasing strongly in adults. With the decrease in adults also came some changes in the dOri-driven response modulation, as evidenced by significant interactions between dOri and age when comparing the P44–47 age group with the adults using a two-way ANOVA (contribution of age, p<0.001. interaction, p=0.03). The overall decrease in V1 plaid responses was very pronounced in our data set. Note, however, that a previous study (Popović et al., 2018) investigating the development of V1 cross-orientation inhibition in ferrets (i.e., looking at responses to plaids with dOri = 90 deg) did not observe changes in plaid responses between P40 and adulthood, something that will need to be resolved with further experiments (see also Discussion).

Figure 5 with 1 supplement see all

Download asset Open asset

Figure 5—source data 1

Neuron-by-neuron metrics for Figure 5 and Figure 5—figure supplement 1.

https://cdn.elifesciences.org/articles/59798/elife-59798-fig5-data1-v2.xlsx

Download elife-59798-fig5-data1-v2.xlsx

Second, we made an unexpected discovery in our V1 data set: In adult animals, V1 was strongly dominated by component responses (see Figure 5E,F and Lempel and Nielsen, 2019). The same held true for the youngest age group: Across the P37–40 animals, we observed a similar number of pattern and component cells as in adults (pattern: 4% versus 3%, component: 87% versus 95%; Figure 5E), and the pattern index distributions for both groups closely aligned (Figure 5F). However, the middle age group deviated from this pattern. Across animals aged P44–47, we observed a larger number of pattern cells (15%), and the pattern index distribution was significantly shifted toward stronger integration (p<0.001). It should be noted that even with this shift V1 pattern indices remained significantly lower on average than PSS pattern indices (Figure 5—figure supplement 1, p<0.001). The same was true also for the two other age groups (p<0.001 for both groups). Taken together, our data suggest a temporary change in V1 plaid responses, characterized by stronger integration of the two component directions and weaker inhibition caused by the simultaneous presence of two gratings. The combination of these two changes had an interesting consequence: V1 responses were enhanced for plaid conditions in which the pattern motion was in the preferred direction, but were left largely unchanged for conditions in which the components moved in the preferred direction (Figure 5A). Thus, increases in response strength were largely reserved for pattern responses.

In an effort to identify possible sources for these changes, we next recorded V1 plaid responses at P44–47 while inactivating PSS with the GABA_A agonist muscimol. Inactivating PSS reversed both temporary changes in V1 plaid responses (Figure 6A,B). The relative plaid response strength was significantly reduced relative to control data (ANOVA normal versus muscimol p<0.001), reaching the same levels seen normally at P37–41. Similarly, PSS inactivation resulted in a significant shift of the pattern index to lower values (p=0.01) that matched both younger and older animals (see Table 1 for statistics). A sham PSS injection without muscimol (Figure 6C,D) did not impact V1 relative plaid responses or pattern indices (see Table 1 for statistics), confirming that PSS inactivation, rather than side effects of the injection, caused the observed V1 effects. Thus, both changes in V1 plaid responses are at least partially dependent on feedback from PSS.

Figure 6

Download asset Open asset

Figure 6—source data 1

Neuron-by-neuron metrics for Figure 6.

https://cdn.elifesciences.org/articles/59798/elife-59798-fig6-data1-v2.xlsx

Download elife-59798-fig6-data1-v2.xlsx

As described in the previous section, we observed temporary changes in V1 plaid responses between P44 and P47. The results of the inactivation experiments indicate that these changes depend on feedback from PSS. At the same time, changes in V1 presumably propagate to PSS (and more generally areas downstream from V1), potentially initiating changes in plaid processing throughout the motion pathway. What consequences could the V1 changes have on motion integration computations in PSS, in particular relative to developmental processes occurring within PSS? Obviously, this question will need to be answered experimentally, but as a first step toward providing some answers we made use of a computational model for the motion pathway.

We have previously shown that a two-stage model consisting of V1-like local direction filters followed by a PSS-like integration stage could fit well to neural responses in PSS (Lempel and Nielsen, 2019). Here, we used a modified version of this model to investigate development. Since the goal was to identify the contribution of particular V1 and PSS changes, we replaced the front end of the model. Rather than adjusting the parameters of the V1 stage as needed for the best fit to a given PSS neuron, we fixed it to match the experimentally observed data. More precisely, in the model implemented here (see Figure 7A), the first stage was set to be a bank of 16 direction filters that when exposed to plaids would reproduce our measured average V1 responses as closely as possible, including their dOri dependence. Two modifications were made to the empirical V1 response profile in building the model V1 filters: First, we made responses symmetric around the preferred direction. Second, motion pathway models usually assume perfect direction tuning in the V1 stage (i.e., no responses to stimuli moving in the null direction). We similarly modified the V1 response profile to achieve maximal direction tuning (see Materials and methods for details).

Figure 7

Download asset Open asset

Figure 7—source data 1

Neuron-by-neuron metrics for Figure 7.

https://cdn.elifesciences.org/articles/59798/elife-59798-fig7-data1-v2.xlsx

Download elife-59798-fig7-data1-v2.xlsx

The PSS integration stage of the model remained identical to our previous study. In this stage, the responses of the 16 V1 direction channels were integrated using a weight function with an excitatory component centered at the preferred direction and an inhibitory component centered at the null direction. Both components were modeled as von Mises functions. The width of the excitatory component, as well as the width and amplitude of the inhibitory component were free model parameters. Below, they are summarized as W_exc and W_inh, which refer to the area under the excitatory and inhibitory component, respectively. The integration stage finally was followed by an exponential non-linearity determined by the free parameter T_PSS. Note that one major purpose of the model was to test whether the observed V1 changes could propagate to PSS (and how they would affect PSS responses). For this reason, the model remained strictly feedforward, despite the above observations that V1 changes depended on feedback from PSS. Future models will need to expand to more complicated models including both feedforward and feedback links to fully explore the dependencies between areas.

As a first modeling exercise, we attempted to identify which PSS-internal mechanisms might explain the developmental changes occurring after P40. To this end, we set the V1 stage to the data measured between either P37–40 or P44–47 and then used a maximum likelihood approach to fit the model to PSS responses measured in the same age range (see Materials and methods). Note that because the V1 data only covered P44–47, we restricted the PSS data set to the same range for the modeling analysis. As illustrated by the example pattern and component cells shown in Figure 7B, the model was well able to fit PSS neurons of different types at both ages. This was further confirmed by a generally high correlation between modeled and measured data for both groups, with little differences between the groups (Figure 7C). In all subsequent analyses, we excluded cells with correlation coefficients below 0.5. For both age groups predicted and measured pattern indices generally matched well (Figure 7D). While the model generally overestimated the absolute value of a neuron’s pattern index, likely because of lower noise levels in the modeled responses, pattern cells remained classified as pattern cells based on the model fits, and component cells remained classified as component cells.

The good model fit for both age groups allowed us to compare model parameters across ages. Cumulative distributions for W_exc, W_inh and T_PSS are plotted for both age groups in Figure 7E. These plots show a trend toward increasing W_exc with age, but the difference between age groups did not reach significance (p=0.06). At the same time, W_inh significantly increased between the two groups (p<0.001). T_PSS, finally, remained unchanged. Thus, the model results suggest that changes in PSS null direction inhibition are a major component of motion integration development.

Having identified the major changes internal to PSS that could be driving motion integration development, we next sought to determine how the temporary V1 changes complemented them. To this end, we used the following stepwise procedure: We first initiated the model with its state at P37–40. In this baseline model, the V1 stage was set to match the empirical data for P37–40, as shown in the first column of Figure 8A. We also set the parameters for the PSS stage to those determined by fitting the model to PSS data from the same age group. One set of parameters was computed for each neuron in the data set (N = 136), allowing us to construct a population of model PSS neurons. The resulting model PSS responses are summarized in Figure 8A by plotting the average plaid response profile for all PSS model neurons (third column), as well as the resulting pattern index distribution (fifth column).

Figure 8 with 1 supplement see all

Download asset Open asset

Figure 8—source data 1

Neuron-by-neuron metrics for Figure 8 and Figure 8—figure supplement 1.

https://cdn.elifesciences.org/articles/59798/elife-59798-fig8-data1-v2.xlsx

Download elife-59798-fig8-data1-v2.xlsx

We then used modifications of the baseline model to assess the impact of different developmental changes. First, we asked what contributions the increase in PSS null direction inhibition made to pattern responses. In this model iteration, we kept the V1 stage unchanged relative to the baseline model, that is, it remained set to mimic responses at P37–40. At the same time, we adjusted W_inh of the PSS stage to match the levels seen at P40–44. To achieve this, we increased the value of the model parameter I, which controls the strength of inhibition (see Materials and methods), so that the new mean W_inh value equaled that determined when fitting the model to data for P40–44 (Figure 8—figure supplement 1; note that we did not adjust K_I since it showed little differences between ages in the original fit). Figure 8B shows the resulting average PSS plaid response profile (third column) and pattern index distribution relative to that of the baseline model (fifth column). It also shows the response difference between the new model iteration and the baseline version as a function of plaid direction and dOri (fourth column). These figures demonstrate that while increasing PSS inhibition resulted in a large overall increase in pattern indices, it came with reduced plaid responses. On the one hand, responses were weakened for plaids in which one of the conditions (but not the integrated plaid) moved in the preferred direction (these make up the arms of the ‘V’-shaped response profile). Changes to these conditions will increase pattern responses by decreasing component responses and are the key mechanism by which inhibition can shape pattern responses. On the other hand, the increased inhibition extended to conditions in which the pattern movement was in the preferred direction, in particular for intermediate dOri values (likely because for these conditions the components were far enough from the preferred direction to experience some of the increased null direction inhibition). Thus, the improvement in pattern index by increased inhibition came at a cost.

In the second model iteration, we assessed the impact of increasing the excitatory integration of component signals in PSS. While the changes in W_exc between P37–40 and P44–47 failed to reach significance, excitation levels nonetheless showed a moderate increase between these two groups. Here, we kept all model parameters with the exception of K_E, the only parameter controlling excitation, fixed in their baseline state. K_E was decreased (which has the effect of broadening the excitatory part of the integration function, thereby increasing W_exc) until the mean of the model W_exc distribution matched that determined previously in the P44–47 model fit (Figure 8—figure supplement 1). Figure 8C again plots the resulting model PSS response to plaids, the difference to the baseline model, and the new pattern index distribution. Increasing excitation actually decreased the pattern index by diminishing differences between pattern and component responses, but as expected, the strength of pattern responses was increased across all dOri values.

In the third model iteration, we then asked what contribution the observed temporary changes in V1 would make to PSS pattern responses. In this iteration, we changed the V1 stage to match the data measured in P44–47 animals, while fixing the PSS model parameters to the values determined for P37–40. The new V1 response profile is plotted in Figure 8D, as is the resulting average PSS response profile, its difference relative to the baseline model, and the resulting pattern index distribution. We found that changing the V1 stage on its own had little consequences on PSS pattern indices, which remained largely unchanged from the P37–40 baseline model. However, there were changes in the PSS response to plaids. The selective increases in V1 pattern responses indeed propagated to PSS and resulted in increased responses for pattern motion for plaids with intermediate dOri values.

Finally, we combined V1 and PSS changes in the fourth model iteration to determine their combined effect (Figure 8E). The major difference between this model iteration, and a full fit of the pathway model to data from P44 to P47 as performed in Figure 7, was the state of the PSS stage. Rather than optimizing these parameters through fitting the model, we only shifted the mean inhibition and excitation levels of the baseline model as described above. In this last model iteration, the increases in inhibition again served to decrease component responses and increase the pattern index. At the same time, increases in PSS excitation, combined with the changes in V1 plaid responses, mitigated the side effects of the increased suppression levels by limiting them to the component responses only.

All procedures adhered to the guidelines of the National Institute of Health and were approved by the Animal Care and Use Committee at Johns Hopkins University. Experiments were performed in both male and female sable ferrets (Mustela putoris furo) aged 37–448 days. Experiments were conducted in anesthetized ferrets, using the same procedures as described in our previous work (Lempel and Nielsen, 2019). Briefly, animals were anesthetized during the experiments using isoflurane (during surgery: 1.5–3%, during recording: 0.5–2%) and paralyzed using pancuronium bromide (0.15 mg/kg/hr). A number of vital parameters (heart rate, SpO2, EKG, EtCO2, and the EEG) were monitored continuously to ensure adequate anesthetic depth during the experiments. Neosynephrine and atropine were applied to the eyes to retract the nictitating membrane and dilate the pupil, and animals were fitted with contact lenses. Before recordings, craniotomies were made above either V1 or the posterior bank of the suprasylvian sulcus to reach PSS. We targeted central visual field regions in V1, and central and more peripheral visual field regions in PSS.

Neural signals were recorded using either custom-made tetrodes (12 µm nichrome wire, gold-plated to reach final impedances of 150–500 kΩ) or 64-channel silicon microprobes obtained from the Masmanides lab at UCLA (Du et al., 2011). Probes were gold-plated to reach final impedances of 150–300 kΩ. Signals were amplified and recorded using a CerePlex Direct amplifier (Blackrock Microsystems) or a RHD2000 amplifier (Intan Technologies). Raw data were acquired at either 30 or 20 kHz and filtered between 250 Hz and 5 kHz. Spike detection threshold was set manually for each recording based on noise levels. Single-unit isolation was performed off-line using MATLAB (MathWorks). Isolation was based on multiple waveform characteristics (e.g., spike amplitude peak, area under the waveform, repolarization phase slope) recorded on the four tetrode channels or on neighboring channels of the silicon probe. Quality of isolation was confirmed by inter-spike interval (ISI) analysis. Units that displayed ISIs below 1.2 ms were not included in further analyses.

PSS was inactivated by injecting muscimol (2.5 mg/ml in ACSF; Sigma) in multiple sites spanning the posterior bank of the suprasylvian sulcus. For control experiments, ACSF without muscimol was injected instead. To aid the visualization of injection location and spread, 0.05% Fast Green FCF (Fisher Scientific) was added to both the muscimol and control solution. To inject the solution, we used glass pipettes pulled using a micropipette puller (Sutter instrument, model P-1000) and 5 µl calibrated glass capillary tubes (VWR). Infusion was performed at a rate of 0.1–0.5 µl/min. For each animal, injections were performed at four different locations along the posterior bank of the suprasylvian sulcus. At each location, 0.2–0.4 µl of either muscimol or ACSF solution were infused at 4 depths spaced 200–300 µm apart. In total, 4–5 µl of either muscimol of ACSF solution were injected per animal. Recordings were made in V1 1–5 hr after injection. For muscimol experiments, lack of PSS responsiveness was confirmed by extracellular recordings either during or right after the last V1 recording.

Visual stimuli were generated using the Psychophysics Toolbox extensions for MATLAB (Brainard, 1997; Pelli, 1997) and displayed on a 24-inch LCD monitor with a refresh rate of 120 Hz, placed 25–35 cm in front of the ferret. For a subset of experiments, a 43-inch LCD monitor with a refresh rate of 60 Hz was used instead. Monitors were gamma corrected using a SpectraScan 655 (PhotoResearch).

Direction selectivity, pattern index, pattern and component Z-corrected correlations were compared between experimental groups using a Welch’s t-test. This version of the Student’s t-test is modified to accommodate unequal sampling and unequal variance across experimental groups. Model parameters determined through data fits were compared between age groups using the non-parametric Wilcoxon rank-sum test.

In cases of significant sample size differences between two groups (e.g., different age groups for PSS data and model fits), we performed an additional re-sampling test to verify significant differences in the median of a given metric (such as the pattern index or a model parameter). One thousand data sets were computed by randomly subsampling the larger of the two groups. In the random re-sampling procedure, N data points were randomly drawn (with replacement) from the larger data set, where N equals the number of samples in the smaller data set. For each of the 1000 data sets, we then computed the median and compared it to the median of the entire (larger) data set by computing their difference. This yielded a distribution of median differences for the 1000 randomly subsampled data sets. Finally, the difference in medians between the two experimental groups was compared against the difference distribution determined by the re-sampling procedure to obtain the test’s p-value.

To test for differences in relative plaid responses across dOri values, age groups, and their interaction, a two-way ANOVA with an interaction term was used. We found that the distribution of relative plaid responses had a long tail at the high end, typical of a χ²-distribution. To transform this distribution in a way that would better fit the assumptions of this parametric test, we performed a Wilson–Hilferty transformation (Wilson and Hilferty, 1931) of the data by computing the cubic root. This transformation largely removed the asymmetries in the distribution (Figure 4—figure supplement 1). This transformation was also used before computing correlations between relative plaid responses and pattern index (Figure 4C). For comparisons of Wilson–Hilferty transformed relative plaid responses at particular dOri values, a Welch’s t-test was used.

The model used here is a modification of a model we described previously (Lempel and Nielsen, 2019). The code for the new model is available on github (https://github.com/nielsenlabmbi/PSSModelMLE; Lempel, 2020; copy archived at swh:1:rev:7a4bba49cd2c72fab0623f3c3756fd55506cc816 Lempel, 2021). The LGN and V1 stages of the model described in Lempel and Nielsen, 2019 were removed and instead responses of V1 cells providing input to PSS were modeled after empirical data collected from V1 neurons at different developmental stages. Model V1 responses were computed from empirical data in the following way: Data from individual neurons (as a function of plaid direction and dOri) were first normalized by their maximum response to gratings and shifted in direction space to align the preferred direction to 0 deg. Data from neurons belonging to the same age group were then averaged to compute a single V1 filter profile per group. This profile was used to generate 16 V1 filters with different direction preferences (homogeneously covering direction space) by shifting the profile in direction space.

In the previous implementation of this model (Lempel and Nielsen, 2019), V1 response profiles were completely direction selective and symmetric around the preferred direction, in that stimuli moving in directions equidistant to the preferred direction elicited identical responses. To maintain these two properties of V1 filters in the new model implementation, we applied the following two transformations: To make the filters symmetric, we effectively mirrored the 2D response profile for each filter around the preferred direction. More precisely, at each dOri the filter response for directions with similar distances from the preferred direction were averaged. The original responses were then replaced with the average. To make the output of the V1 stage perfectly direction selective, we combined the responses of V1 model neurons with opposite direction preferences. More precisely, each direction channel consisted of a pair of V1 filters. The response of the filter was computed as P-P(o)O, where P is the response of filter one at a particular direction and dOri, O the response of filter two which prefers the opposite direction, and P(o) the response of filter one to a grating moving in the null direction. This forced the response of the filter pair to be 0 at the null direction. Any negative values resulting from the computation were set to 0.

The V1 filter stage was followed by a PSS integration stage as described previously. Briefly, the responses of the 16 V1 filters were combined linearly, using a weight function that consisted of two van Mises functions, one centered at the preferred and one centered at the null direction. This PSS integration stage was followed by a final non-linearity, implemented here as an exponential function (as opposed to the subtraction and rectification method implemented previously) so that:

This change was implemented to aid the parameter optimization method described below. In total, the model had the following free parameters: K_E, K_I, I, T_PSS, and Resp_max. This last variable simply multiplied PSS responses computed by the model to best match empirically recorded firing rates.

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

We are grateful to J Killebrew, W Nash, and W Quinlan for technical support and R Krauzlis for advice on the muscimol injections. We thank R Srinath, E Dunn-Weiss, A Emonds, and other members of the Nielsen and Connor labs for helpful discussions and experimental support.

Animal experimentation: All procedures were performed in strict accordance with the guidelines of the National Institute of Health and were approved by the Animal Care and Use Committee of Johns Hopkins University (Protocol Numbers FE15M305 and FE18M239).

1. Received: June 8, 2020
2. Accepted: March 8, 2021
3. Version of Record published: March 22, 2021 (version 1)
4. Version of Record updated: March 24, 2021 (version 2)

© 2021, Lempel and Nielsen

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.