Spatiotemporal input reorganization and enhancement of input-output gain sharpen cortical direction selectivity during development

In the primary visual cortex (V1) of carnivores and primates, early visual experience shapes the development of functional properties such as selectivity for direction of motion. However, it remains unclear which aspects of the cortical circuitry - synaptic and/or cell-intrinsic - are molded by the visual activity, and how. Therefore, we performed in vivo intracellular recordings from V1 simple cells in visually naive and experienced ferrets, and computed their membrane potential (Vm) spatiotemporal receptive fields and Vm-to-spike input-output transfer functions. Comparison across the two developmental stages revealed marked reorganization of inputs and a sharp enhancement of input-output gain in the experienced state, leading to prodigious amplification of direction selectivity at the spiking level. Simultaneously, we detected a lowering of spike thresholds and an intensification of gamma-band Vm oscillations, which might explain the enhancement of spiking. Thus, direction selectivity develops through combined processes of synaptic re-wiring and maturation of cell-intrinsic properties.


Introduction
The primary sensory cortices construct a representation of the physical world by integrating neural response selectivities for distinct sensory features. The receptive fields underlying these feature selectivities arise through a combination of the precise patterns of synaptic inputs received by cortical neurons and their intrinsic cellular properties. A cortical neuron receives thousands of cortical and sub-cortical synaptic inputs and generates a spiking output that reflects its response selectivity. The cell-intrinsic properties gate the input-output transformation by controlling how the neuron summates its inputs and then maps the summed input to a spiking output. Determining how the synaptic input patterns and the cell-intrinsic properties of cortical neurons evolve during development is a key component to understanding how a mature cortical circuit performs computations in general.
In the primary visual cortex (V1), neuronal responses are characterized by selectivity for two features of visual stimuli: orientation (orientation selectivity, OS) and direction of motion (direction selectivity, DS) [1][2][3] . In ferrets, a small carnivore, OS develops early through innatelyspecified processes that do not require visual experience, and consequently V1 neurons exhibit significant OS already at the time of eye-opening [4][5][6] . In contrast, DS -defined as a neuron's preference for one of the two directions of motion orthogonal to its preferred axis of orientationtypically develops over a period of 1-2 weeks after eye opening through a process that requires visual experience, and does not form in dark-reared or strobe-reared animals 4,5,[7][8][9] . In directionselective simple cells of V1 in visually experienced animals, the spatially selective synaptic inputs activate the cell with varying latencies in a specific temporal order [10][11][12] . These subthreshold inputs are summed linearly, thereby generating direction-selective responses at the subthreshold level, and are then passed through a threshold non-linearity to obtain a direction-selective spiking response [13][14][15] . During development, an increase in DS at the level of spiking could be brought about by an improved spatiotemporal summation 16,17 , an enhancement of inputoutput transfer non-linearity 13 , or both. During the first weeks of vision, visually-driven neural activity must transform the patterns of synaptic connectivity and the cellular properties of simple cells to instantiate these changes and lead to enhanced DS. To understand this process, it is necessary to study the evolution of the input patterns and the input-output transformations of simple cells over the period when DS emerges.
Therefore, we carried out in vivo sharp intracellular recordings from V1 simple cells in visually naïve and experienced ferrets and compared the spatiotemporal patterns of synaptic inputs as well as the input-output transfer functions across these developmental stages.
Comparison of the simple cell spatiotemporal receptive fields (STRFs) between naïve and experienced ferrets revealed that there were significant reorganizations in the spatiotemporal patterns of inputs in the experienced animals, leading to more optimal linear summation of inputs and consequently to a modest increase in Vm responses to the preferred direction of motion.
Further, the slope of the non-linear Vm-to-spike transfer function underwent sharp enhancement during the same period of development. This enhancement in gain was brought about by lowering of spike threshold and increase in high frequency oscillations of Vm in the experienced animals. As a result of these modifications, the modest improvements in Vm responses induced by STRF rearrangements mapped on to a steeper input-output transfer curve in the experienced animals, leading to a sharp enhancement in DS at the spiking level.

Results
To study the subthreshold mechanisms of development of direction selectivity (DS) in the primary visual cortex (V1) of ferrets, we carried out in vivo intracellular recordings in V1 using sharp microelectrodes. Experimental animals were split in to two age groups: visually naïve ("naïve": age P30-34, n = 10 animals, 23 cells) and visually experienced ("experienced": age P40-60, n = 11 animals, 29 cells). Cells were recorded under isoflurane anesthesia while visual stimuli were displayed on a computer monitor. Sinusoidal gratings at varying orientation angles (in steps of 30 degrees) drifting in either of the two directions orthogonal to the axis of orientation were used to map orientation and direction tuning of neurons. The degree of orientation and direction selectivity were expressed using orientation selectivity index (OSI) and direction selectivity index (DSI), respectively, both of which varied between 0 to 1 (see Supplementary methods). Because V1 simple cells have a substantial linear response 15 , Vm and spike responses of these neurons to drifting grating stimuli are periodically modulated following the temporal frequency of the grating 18 . Simple cells are typically distinguished from complex cells by comparing the fundamental (F1) and the DC components (F0) of the responses 14,18,19 .
We first converted the Vm and spike responses of each recorded neuron to the frequency domain using Fast Fourier Transform (FFT) and then analyzed all response properties for the F0 and F1 components separately as well as combined (F0+F1). We calculated Vm and spike modulation ratios by dividing the amplitude of the F1 component by that of the F0 component of the respective responses. Cells with spike modulation ratio equal to or greater than 1 were classified as simple, while the rest were categorized as complex (see Supplementary Figure S1). Neurons that did not fire sufficient spikes were classified as simple if their Vm modulation ratio was greater than 0.5. According to this classification, our data set contained 26 simple cells (10 in naïve and 16 in experienced groups) and 26 complex cells (13 each in naïve and experienced groups). In the following sections, we first describe the various analyses performed on the simple cells, while relevant data from the complex cells are provided in supplementary figures S3, S7-9.

Emergence of Direction Selectivity following visual experience
We saw a sharp rise in visually-evoked spiking responses from both simple and complex cells of the experienced animals. The increased responsivity could be detected in two ways. First, the fraction of cells that fired action potentials in response to drifting grating stimuli was significantly higher in the experienced group (simple: naïve, 7/10; experienced, 16/16; p < 0.05, complex: naïve, 7/13; experienced, 13/13; p < 0.05, Fisher Exact test). Thus, every recorded cell in the experienced group fired action potentials, but 9/23 cells in the naïve group did not fire any action potentials at all, even though subthreshold visual responses could be detected in these neurons. Second, the magnitude of spiking responses to the preferred direction of motion was significantly stronger in the experienced group (Supplementary Figure S2; naïve vs experienced, for both simple and complex cells, p = 0.001, Wilcoxon Rank Sum (WRS) test). Therefore, spiking responses dramatically increased in the visually experienced animals.
Direction selectivity, measured both at the sub-and suprathreshold levels, increased in the experienced animals. Figure 1A shows an example simple cell each from a naïve and an experienced ferret. The neuron from the naïve ferret ( Figure 1A, left panel) responded strongly to horizontal gratings, irrespective of whether they were moving upward or downward, while showing only weak subthreshold responses to vertical gratings. In contrast, the neuron from the experienced ferret ( Figure 1A, right panel) strongly responded only to vertical gratings drifting to the right while responding minimally to all other stimuli. Thus, while both neurons exhibited orientation selectivity, only the experienced neuron exhibited direction selectivity as well. To describe the feature selectivity of these neurons, we plotted their F0 and F1 response magnitudes for each direction of motion in the form of tuning curves. In the tuning curves of the naïve neuron ( Figure 1B, left panel), two strong peaks in the direction space separated by ~180 degrees signified existence of orientation selectivity along with low direction selectivity. In contrast, for the experienced neuron ( Figure 1B, right panel), one strong peak in the direction space signified both high orientation and direction selectivity.
To quantify the degree of direction selectivity, we calculated a direction selectivity index In a majority of simple cells, the DSI for spiking responses were higher than the DSI for Vm responses. A pairwise comparison of Vm and spike DSI within each neuron revealed that the spike DSI was significantly higher than Vm DSI in both the naïve and the experienced groups ( Figure 1D; naïve, p = 0.04, n = 7; experienced, p < 0.001, n = 16; signed rank test). Therefore, in both groups of animals, the outputs of V1 simple cells were more sharply tuned for direction than their summed inputs, hinting at non-linear transformation from Vm inputs to spiking outputs 13,18 .
It is evident from the individual responses and the tuning curves in Figure 1A and 1B that, in addition to the linear response component (fundamental or F1) that modulates following the stimulus contrast, the responses of simple cells also contain a significant non-linear component (DC or F0: the elevation of Vm from the baseline floor). To test if the developmental enhancement of DSI in the simple cells were restricted to one of these two components or affected both, we calculated DSI separately for the F0 and F1 response components. An increase in DSI could be realized either by an increase in responses to the preferred direction, a decrease in responses to the null direction, or both. In order to delineate the response dynamics underlying the enhancement of DSI, we analyzed the Vm and spiking response amplitudes to the preferred and the null direction stimuli for each simple cell. The total (F0+F1) responses to the preferred and null stimuli were compared between the naïve and experienced groups, separately for Vm and spiking responses (Figure 1F). At the level of Vm (Figure 1F, top), the null responses (left) do not show any significant change over development, whereas the preferred responses (right) show a modest but significant increase from the naïve to the experienced group (null, naïve: mean = 10.9 +/-1.4, n = 10; experienced: mean = 10.6 +/-1.05, n = 16, p = 0.8, WRS test; preferred, naïve: mean = 12.7 +/-1.3, n = 10; experienced: mean = 17.1 +/-1.2, n = 16, p = 0.02, WRS test). At the level of spiking ( Figure 1F, bottom), a similar pattern was observed (null, naïve: mean = 4.9 +/-1.9, n = 10; experienced: mean = 8.3 +/-2, n = 16, p = 0.15, WRS test; preferred, naïve: mean = 7.01 +/-2.4, n = 10; experienced: mean = 28.6 +/-4.2, n = 16, p < 0.001, WRS test). Therefore, the increase in DSI following visual experience was driven by an increase in responses to the preferred direction of motion and not by a decrease in response to the null direction of motion. Notably, while this increase was modest (35%) at the Vm level, at the spiking level the enhancement of responses to the preferred direction was dramatic (302%).
In summary, visual experience during the 2-4 weeks following eye opening led to a modest increase in Vm responses to the preferred direction of motion, but, this modest increase in Vm responses led to a much stronger response enhancement at the spiking level, thus leading to a robust increase in spike DSI.
It is unclear whether this large enhancement of spike DSI is simply a reflection of the modest enhancement of Vm DSI alone, or whether other mechanisms were involved in amplifying the sub-threshold DSI enhancement. An improved spatio-temporal summation of sub-threshold inputs could account for the observed Vm DSI enhancement. While this alone might fully explain the spike DSI enhancement, other mechanisms such as developmental increase in inputoutput gain could also further amplify the modest Vm DSI enhancements and lead to greater spike DSI. To parse out these possibilities, we carried out two sets of analyses. First, to compare spatiotemporal summation of Vm between naïve and experienced groups, we used reverse correlation of Vm to sparse white noise to map out the Vm spatio-temporal receptive fields (STRFs) 11,12,[20][21][22] of the simple cells in our data set. Second, we constructed the input-output transfer functions of all V1 simple cells by comparing their average spike outputs to the average Vm levels 13,14 , and tested if the input-output gain underwent developmental enhancements.

Mapping spatio-temporal receptive fields of simple cells
Intracellular recordings revealed that the Vm responses became more direction-selective in the visually experienced animals ( Figure 1BCEF). In simple cells, direction selectivity is thought to arise from spatially selective inputs activating the cell at temporally varying latencies 10,15,16,23 . According to this model 24,25 , a direction selective neuron is maximally activated when its inputs selective for specific spatial locations within the receptive field are activated by a moving stimulus in a specific temporal order -the inputs with the longest latencies activating first and inputs with progressively shorter latencies sequentially activating thereafter. This spatiotemporally sequenced order of activation ensures that the subthreshold inputs activated by the preferred direction of motion arrive at the soma simultaneously and achieve maximal summation, leading to a strong response. In contrast, a stimulus moving in the null direction activates the shortest latencies first followed by progressively longer latencies, thereby leading to suboptimal summation of activity, and consequently lower responses in the postsynaptic neuron.
A direction-selective simple cell with its inputs organized in this manner exhibits a slant in its spatio-temporal receptive field (STRF). Could the lower Vm responses detected in visually naïve ferrets be a consequence of immature, non-slanted STRF structures? To test this idea, we computed the Vm STRFs in a subset of V1 simple cells from our data set (naïve: n = 8, experienced: n = 12), and compared various shape parameters describing the STRF structure.
Once a simple cell's preferred orientation was calculated from its tuning curve ( Figure 1B), linear STRFs were computed by reverse correlating the neuron's spike-filtered Vm responses to a sparse 1-dimensional white noise stimulus 13,21,22 . The noise stimulus consisted of black, white and gray bars, angled at the cell's preferred orientation, that changed pattern every 100 milliseconds (see Supplementary methods, Supplementary Figure S4). Therefore, for every spatial location on the monitor, the contrast trace varied between 1, 0 and 1 over time, and this trace was cross-correlated with the spike-filtered Vm trace recorded during the stimulus. High cross-correlation values were obtained if a white bar (contrast 1) led to increased Vm or a black bar (contrast -1) led to decreased Vm (ON subunit), and low cross-correlation values were obtained if a white bar led to decreased Vm or a black bar led to increased Vm (OFF subunit).
For every spatial location the cross-correlation values were plotted at varying lag times, thereby allowing assessment of the latencies at which the high or low cross-correlation values were obtained. Because visually driven spiking activity was weak in naïve ferrets, we focused our analysis primarily on the Vm responses, for which a reliable STRF could be computed for every simple cell tested. Figure 2A shows Vm STRFs of 2 cells recorded from a naïve ferret (cell # 102 and 121) and 2 cells recorded from a experienced ferret (cell # 60 and 80). Consistent with previous findings [11][12][13]16 , the STRF subunits from the high-DSI neurons in experienced animals were slanted in space-time. For cell # 60, gratings moving from right to left on the monitor would activate the inputs with longer latencies first, progressively moving to shorter latencies as the grating moves across the visual field, thereby leading to maximal response. Stimulus moving from left to right, on the other hand, will activate the shorter latency inputs followed by the longer latency inputs, thereby leading to sub-optimal responses. The STRF of the Cell # 80 was also slanted, albeit in the opposite direction, thus signaling that this cell would prefer gratings moving from left to right on the monitor. Indeed, the preferred directions obtained from the tuning curves of these cells matched the direction preference predicted by the STRFs. The STRFs from the low-DSI neurons in naïve animals were not completely unstructured and exhibited several experienced properties such as alternating ON and OFF subunits and a small slant in space-time. However, the STRF subunits from the experienced cells were longer and narrower in profile, and relatively more slanted in space-time.
To rigorously quantify the degree to which STRF subunit structure differed between naïve and experienced animals, we fitted ellipses to the individual ON and OFF subunits 26  parameters were significantly different between the naïve and the experienced groups: eccentricity, major axis, minor axis, orientation, temporal extent and minimum latency. The shape parameter that is most relevant to direction selectivity is the eccentricity of the ellipses describing the STRF subunits. If STRF subunits are broad along the minor axis of the ellipse, they will enable some response summation from stimuli moving in either direction, thus leading to weaker direction selectivity. Also, if STRF subunits are short along the major axis of the ellipse, there might not exist sufficient latency differences to facilitate differential input summation. Therefore, to optimally support high direction selectivity, STRF subunits would have to be narrow on the minor axis and elongated on the major axis, which would result in high eccentricity. Consistent with this idea, we found that the eccentricity measured from the experienced subunits were significantly higher (p = 0.001), and was caused by both a longer major axis (p = 0.03) and a shorter minor axis (p = 0.03). Also, the subunits from the experienced animals subtended a smaller angle on the space axis (orientation: p = 0.005), which would predict the cells from experienced animals to be tuned to higher motion velocities, consistent with data from actual V1 recordings in cats 17 . The overall area (p = 0.6) and the spatial projection (p = 0.35) of the subunits did not significantly differ. However, the projection on the latency axis was significantly longer in the experienced group (p = 0.04), suggesting that a longer range of response latencies became available in the experienced animals. This extension of the latency range was achieved solely by stretching the STRF subunits towards the lower latency side: while the maximum latencies of the subunits were not significantly different (p = 0.3), the minimum latencies were significantly lower in the experienced group (p = 0.05). This suggests recruitment of lower latency inputs in the experienced animals 27 . Consistent with this view, we found that the Vm response latency distributions were significantly shifted towards lower latencies in the experienced animals (Supplementary Figure S5).
If the changes in Vm STRF structure resulted in improved discrimination between preferred and null directions of motion, the STRF structure parameters should correlate with Vm DSI measured in each cell. Therefore, we computed linear correlations between cell-averaged shape parameters and the cells' Vm DSI values ( Figure 2C). We found that the Vm DSI of the cells increased with increasing 'narrowness' of the subunits (longer 'major axis', shorter 'minor axis' and higher 'eccentricity') and decreasing minimum latencies accessible (lower 'min latency').
Other parameters did not show significant correlation. These data further corroborate the idea that narrower spatio-temporal input profiles and availability of shorter-latency inputs exert a strong impact on directional summation of inputs in simple cells of experienced animals, thereby leading to higher direction selectivity.
We tested the linearity of spatiotemporal input summation in the simple cells in two ways.  Figure S6B). The average R 2 per cell for linear correlation between predicted and actual Vm traces were indistinguishable between naïve and experienced groups and the overall R 2 was 0.37. Thus, according to both methods, about a third of the variance in Vm dynamics could be explained by the linear STRFs, leaving the rest of the variance to non-linear and stochastic processes. But this sizable non-linear component was not distinguishable between naïve and experienced groups, thereby ruling out altered non-linearity of sub-threshold summation as a potential mechanism by which Vm DSI is enhanced in the experienced group.
In summary, following visual experience, the Vm receptive fields exhibit marked reorganization involving narrowing of the space-time profile and recruitment of lower latency inputs, leading to improved discrimination of directional inputs at subthreshold level and thus a more direction-selective Vm response. In order to get a complete picture of the developmental change in V-F relationship, we pooled the individual V-F curves from each cell within a category (naïve or experienced), binned them at 1mV intervals, and fitted the hyperbolic tangent function to these pooled data points. As a result, we obtained a single population average V-F curve for the naïve and the experienced groups each (Figure 3E; for complex cells, see Supplementary Figure S7). The population V-F plots were fitted by the following hyperbolic tangent functions: naïve: FR = 20.8 + 21.4 * tanh((Vm -19.1)/8.9)); experienced: FR = 21 + 20.3 * tanh((Vm -13.9)/4.5)). These plots capture the essence of the result -the slope of the V-F relationship in the experienced group is much higher than in the naïve group, signifying a strong input-output gain enhancement with development.
To test the validity of our conjecture that enhanced V-F gain led to increased firing in the experienced cells, we projected the average Vm responses to preferred and null direction stimuli recorded in each cell on to the V-F plot fits from the same cells and predicted average firing responses. Comparison of the predicted and measured spiking responses revealed a strong significant correlation between the two, in both the naïve and experienced groups ( Figure 3F; top: naïve, preferred direction: R 2 = 0.93, p < 0.001, null direction: R 2 = 0.84, p = 0.001; bottom: experienced, preferred direction: R 2 = 0.78, p < 0.001, null direction: R 2 = 0.77, p < 0.001).
These strong correlations validate that our fitting of the V-F plots captured closely the essence of the cells' V-F relationships.
We earlier noted that, in the experienced animals, a modest increase in Vm responses to the preferred direction of motion led to a much stronger response at the spiking level (Figure 1F), leading to a robust developmental increase in spike DSI. Detection of a strong enhancement of input-output gain raised the possibility that this enhancement amplified the modest gains in Vm responses at the spiking level. To test this idea, we took the average total (i.e., F0+F1) Vm responses to preferred and null directions for each cell, and projected them on to both the naïve and the experienced population-average V-F fitted plots (Figure 3E), thereby generating spiking rate estimates for 4 conditions: naïve Vm projected on to naïve V-F, experienced Vm projected on to naive V-F, naive Vm projected on to experienced V-F and experienced Vm projected on to experienced V-F. Among these, the naïve-on-experienced and experienced -on-naïve conditions allowed us to estimate what levels of spiking would be obtained if only input-output gain or Vm response gain had taken place, respectively. From these spiking estimates we calculated predicted spike DSI for each category. A comparison of these 4 categories of predicted spike DSI with the actually measured spike DSI in naïve and experienced animals ( Figure 3G) revealed that only when experienced Vm was projected on to the experienced V-F plot, we obtained spike DSI predictions that were indistinguishable from the actually measured spike DSI in the experienced animals (one-way ANOVA, F = 3.64, p = 0.006; Fisher pot-hoc test p-value > 0.05 only for the experienced -on-experienced vs measured-experienced conditions). The predicted spike DSI for naïve-on-experienced and experienced -on-naïve conditions were both significantly lower than the measured experienced spike DSI (Fisher pot-hoc test p-value < 0.05). Therefore, an increase in Vm responses to the preferred direction of motion, presumably effected by the spatio-temporal input reorganizations (Figure 2), and an enhancement of inputoutput gain (Figure 3) were both required to account for the full extent of spike DSI increase during visual development.

Mechanisms of input-output gain enhancement: lowered spike threshold
What mechanisms underlie the sharp enhancement in input-output gain in the visually experienced animals? One idea is that the biophysical spike threshold of the neurons could be lowered during development (Figure 4A, left). The membrane potential undergoes spontaneous and stimulus-driven fluctuations, some of which takes it past the threshold voltage resulting in firing of action potentials. If the threshold voltage were lowered, the same Vm trace will now undergo more threshold-crossings and therefore lead to more spiking on average. Indeed, from analysis of V-F curves we learned that the average voltage at which the average spiking rate of the neurons started to rise above zero -the Vm intercept -was lower in cells from the experienced animals (Figure 3D), suggesting that the threshold voltage on average might be lower in these animals. However, a more rigorous assessment of the threshold voltage for every single spike could be obtained by analyzing the shape of individual spikes. We employed a widely-used algorithm for such shape-based assessment of spike threshold [28][29][30] . Figure 4B shows the Vm trace of a neuron recorded during presentation of visual stimulus. The thresholds calculated for every single spike (see Supplementary methods) is superimposed on the corresponding spikes as red stars. For each simple cell included in the V-F analysis, we collected all spikes during presentation of the preferred-orientation stimulus, measured their threshold voltages, and averaged them to obtain a grand average spike threshold per cell. Comparing these cell-average values across the naïve and the experienced groups, we found a 20% reduction in spike threshold in the experienced animals ( Figure 4C; naïve: 20.14 +/-1.32 mV, n = 8; experienced: 16.05 +/-0.9 mV, n = 15; p = 0.03, WRS test). These results, taken together with the fit-based assessment of threshold voltage (Figure 3D), strongly suggest that biophysical changes during visual development lower spike thresholds in simple cells, thereby allowing them to fire more spikes at a given Vm level.
The slope of the membrane potential (dVm/dt) leading up to a spike is a reliable predictor of voltage-gated sodium current during spike upstroke, and therefore might control spike threshold 29,30 . If a high fraction of the voltage-gated sodium channels exists in the inactivated state, the sodium current decreases, thereby lowering the slope of rise of Vm (dVm/dt). As a result, the maximum dVm/dt reached before the onset of a spike is reduced, and the spike threshold is increased, as has been demonstrated in the cat V1 29,30 . To test if the lowered spike threshold observed in the visually experienced animals could have been caused by increased maximum dVm/dt before spike onset, we calculated the maximum slope of membrane potential in a 10-msec window before each action potential initiation, and computed an average maximum dVm/dt value for each cell. When these values were compared across naïve and experienced animals, a 29% increase in maximum dVm/dt was detected in the experienced group (Figure 4D; naïve: 200.7 +/-15.8 volts/sec, n = 8, experienced: 257.8 +/-13.4 volts/sec, n = 15; p = 0.01, WRS test). Therefore, in visually experienced animals, a steeper rise in Vm before an action potential led to lowering of spike threshold, thereby contributing to increased levels of firing for a given level of Vm.
Interestingly, we detected a similar developmental change in spike threshold and max dVm/dt in complex cells (see Supplementary Figure S8).

Mechanisms of gain enhancement: increased Vm oscillations
Like reduced spike threshold, increased variability or oscillations in the membrane potential could also lead to more threshold crossings (Figure 4A, right) and therefore higher spiking rates 29,[31][32][33][34] . Membrane potential fluctuations, either during stimulus presentation or during intertrial intervals, usually contain energy at various frequency bands. Amongst these, the high frequency (30-100Hz) gamma oscillations are particularly attractive candidate because i) highly synchronized synaptic inputs during visual stimulation could result in high frequency oscillations in Vm, and ii) more frequent or variable rises in Vm are more likely to cause spiking [32][33][34] .
Therefore, to test if the neurons in the experienced animals contained stronger gamma oscillations in Vm, we carried out Fourier analysis of the spike-filtered Vm traces during presentation of stimulus at the preferred orientation (see Supplementary Methods). Figure 5A shows the power in Vm oscillations in the 1-100Hz range for all simple cells. It is evident from the figure that the Vm contained higher power in the gamma frequencies during stimulus presentation in the experienced animals. To quantify the gamma power, we integrated the Vm power over the entire gamma frequency range (30-100Hz) for each individual trial, and averaged the integrated gamma power for all trials within each cell to obtain an average integrated gamma power per cell. Compared to the naïve group, integrated gamma power per cell was significantly higher in the experienced group (Figure 5B; p = 0.016, WRS test). Furthermore, consistent with the idea that higher gamma oscillations in Vm during visual stimulation could lead to higher levels of spiking, we found a strong linear correlation between Vm gamma power and spike rates, both at the levels of single trials and single cells ( Figure 5C). This correlation was significant in both the naïve and experienced groups (per trial: naïve: R 2 = 0.74, p < 0.001; experienced: R 2 = 0.41, p < 0.001; per cell: naïve: R 2 = 0.77, p = 0.002; experienced: R 2 = 0.5, p = 0.002). However, Vm gamma power in cells from experienced animals often reached very high values during many trials, thereby giving a bigger boost to overall spiking in the experienced group. A small but significant increase in gamma power was also observed in complex cells Figure S9). Therefore, in addition to reduced spike thresholds, increased highfrequency oscillations in Vm could also have contributed to increased levels of spiking in the visually experienced animals.

Discussion
We studied the evolution of input-output transformation during visual development of V1 simple cells by mapping out input patterns and neuronal input-output functions through in vivo intracellular recordings. Shape analysis of linear spatio-temporal receptive fields (STRFs) of simple cells in visually naïve and experienced ferrets showed that the elliptical STRF subunits in experienced animals exhibit more eccentric profiles and contain inputs with shorter latencies not seen in naïve animals. This synaptic reorganization led to improved summation of subthreshold inputs, resulting in a small but significant increase in direction selectivity at the level of Vm.
Analysis of the Vm-to-spiking transfer relationship in the same neurons revealed that the slope of the Vm-spike relationship increased in the experienced animals, signifying that for the same average Vm levels in naïve and experienced animals, there was more spiking output in the experienced state. The full extent of DSI increases at the spiking level could only be explained by considering both the small enhancement in Vm direction selectivity and the enhancement of input-output gain. We explored two possible mechanisms for enhancement of spiking in experienced neurons, namely, lowering of spike threshold and increased Vm variability via highfrequency Vm oscillations. Spike thresholds were found to be on 4 mV lower on average in the experienced animals, and there was a significant increase in power in the gamma-band of Vm oscillations, implying that both mechanisms contribute to increased spiking in experienced neurons. Taken together, these results outline a scenario for direction selectivity development that includes two critical components: first, rearrangement in spatio-temporal patterns of inputs to simple cells produces slightly more directionally-selective Vm responses, and second, these modestly enhanced Vm responses then map on to a steeper Vm-to-spike input-output curve, thereby producing an even more directionally-selective spiking response.
In the linear receptive fields of simple cells, optimal directional computation is achieved by making the spatial and temporal properties inseparable, meaning that strong synaptic inputs exist only at certain combinations of spatial position and temporal latencies and that they vary systematically across the space-time axes. STRF mapping in simple cells with low direction selectivity of visually naïve ferrets revealed weaker inseparability, leading to broader STRF subunits that contained inputs with a wider range of spatial and temporal combinations.
However, the naïve STRFs were not completely devoid of structure, and displayed inputs wellorganized into a coherent structure, including even a weak slant in space-time which may explain the small direction selectivity observed at the Vm level in these cells. Over visual development, the set of strong inputs winnows down to a narrower range of space-time combinations, presumably through synaptic pruning, consequently giving rise to more eccentric STRF subunits.
This more eccentric space-time organization of inputs allows optimal input summation for the preferred direction of motion only and thus leads to enhanced Vm DSI.
The non-linear relationship between Vm and spiking in V1 simple cells is well-established.
This results in higher spike DSI compared to Vm DSI in individual cells, as we observed in our data. However, during development, the gain of this non-linearity could increase, further enhancing DSI at the spiking level. Our examination of the Vm-to-spike transfer functions across visually naïve and experienced stages of development corroborated this idea: for increasing average Vm levels, spiking increased with a steeper slope in the experienced animals. Gain enhancement implied that the relatively larger Vm levels attained during the preferred direction of motion mapped on to a much higher level of spiking in the experienced animals, thereby amplifying spike DSI. By cross-projecting naïve and experienced Vm responses on to the naïve and experienced V-F plots, we found that the developmental enhancements of subthreshold responses alone are not sufficient to account for the full range of increase in spike DSI; instead, both subthreshold increases and enhancement of input-output gain must be considered. These results imply that circuit-wide synaptic reorganizations must work in sync with dynamics of cellintrinsic properties during development to give rise to mature functional properties such as direction selectivity.
Although the composition of voltage-gated ionic channels in the cell membrane, specifically in the axon initial segment (AIS), and the location of the AIS with respect to the soma, dictate a neuron's biophysical threshold potential for firing spikes [35][36][37] , the threshold voltage can dynamically vary depending on the manner in which synaptic inputs modulate the neuron's membrane potential 29,30 . For example, convergence of synchronized excitatory inputs, as induced by a sensory stimulus, could produce a sharp rise, and therefore a steeper slope, in the membrane potential, leading to lowering of spike threshold. This contrasts with baseline conditions, when temporally disorganized spontaneous inputs can elevate the Vm to similar average levels but with a shallower slope, and therefore not achieve a lowered spike threshold. This possibility of dynamic regulation of spike threshold makes it an attractive mechanism for controlling a neuron's excitability. Our observation that the spike threshold in the experienced ferrets were ~20% lower relative to rest than in the naïve ferrets makes a strong case for threshold dynamics playing a key role in enhancing the input-output gain in the experienced state. Furthermore, the increased maximum slope of the Vm preceding a spike in the experienced state also suggests enhancement of sodium current during the spike upstroke as a possible reason for lowering of threshold voltage 29,30 . Sensory activity has been shown to drive changes in the location of and the composition of voltage-gated sodium channel subunits in the AIS 38,39 , and such channel dynamics can cause sodium current to increase in the experienced state. Future studies will address if and how visual activity might modulate such channel dynamics and consequently control intrinsic excitability of simple cells.
Altered Vm oscillations in the gamma band has been implicated in state-dependent modifications of firing levels in cortical neurons 29,31,40 . We found a notable increase in Vm gamma power in the visually experienced animals. Because gamma oscillations are shaped by state-dependent changes in network activity, and brain states could vary between stimulus trials, trial-to-trial variation in firing rates could be driven by trial-to-trial variation in Vm gamma power. Consistent with this idea, we found evidence for a strong linear correlation between Vm gamma power and firing rate for each stimulus trial, in both naïve and experienced animals. This relationship held true even at the single cell level. Combined with the observation that, on average, Vm gamma power values were higher in the experienced animals as compared to the naïve animals, these data suggest that increased high-frequency Vm oscillations could constitute one plausible explanation for enhanced spiking activity in the experienced animals.
Strengthening of the cortical inhibitory network and the neuromodulatory circuits during development could underlie the increased gamma oscillations in the network, and will be the focus of future studies exploring the mechanistic underpinnings of input-output gain enhancement.

Methods Summary
Ferrets were anesthetized with ketamine and isoflurane (2% for surgery, 0.08-2% during imaging). Intracellular recordings from V1 neurons were performed using sharp microelectrodes    experienced (bottom, cell # 60, 80) animals. X-axis represents spatial location, y-axis represents latency from onset of visual stimulus, and the cross-correlation coefficient between the stimulus contrast and Vm values are represented by the color. Black lines outline the ellipses fitted to the ON (continuous) and OFF (dashed) subunits. B. Quantification of the 9 parameters defining the characteristics of the STRF subunits, organized in 3 rows and 3 columns. In each panel, the parameter being quantified is described by a schematic on the left, and on the right is a bar plot showing the mean +/-SEM of the parameter values in the naïve (N, green, 19 subunits from 8 cells) and experienced (E, purple, 26 subunits from 12 cells) animals. Black circles denote individual subunit values. Red stars denote statistical significance at p < 0.05 level via WRS test. C. Relationship between cell-average STRF structure parameters and Vm DSI for all simple cells. The R 2 and p values for each linear correlation are shown above each plot.    For the same average level of Vm, spike rate could be increased either by lowering the spike threshold (left) or by increasing the high-frequency oscillations in Vm (right). In the first scenario, the same Vm trace (black) would cross threshold voltage more often if the spike threshold in the experienced cells (purple line) were lower compared to that in the naïve cells (green line). In the second scenario, if the Vm (black) undergoes more high-frequency oscillations in the experienced animals while keeping the threshold (green line) same, increased threshold crossings can also be achieved. B. An example raw Vm trace during stimulus presentation (gray arrows). The calculated spike threshold voltages for every spike in the trace shown as red asterisks, demonstrating the variability in spike thresholds. The inset shows a zoomed in view of one spike marked by the black asterisk and its threshold voltage precisely at the point where Vm rises sharply. C. Comparison of the threshold voltages. Bars represent mean +/-SEM thresholds in naïve (green) and experienced (purple) groups. Each circle shows individual cell values. D. Comparison of the maximum slope of Vm preceding spikes. Bars represent mean +/-SEM max dV/dt values in naïve (green) and experienced (purple) groups. Each circle shows individual cell values.  Top: the power in Vm oscillations are plotted for the entire frequency range of 1-100Hz for all simple cells, each row representing single trials and the trials from the naïve animals arranged on the top (green bar on left side). While at the lower frequencies (1-30Hz, containing theta and beta bands) the power is dominated by the oscillations at the stimulus temporal frequency and its harmonics, in gamma band (30-100Hz) the power of Vm fluctuations can be seen un-corrupted by stimulus-locked oscillations. Bottom: average power at each oscillation frequency, obtained by averaging across trials in each group. B. The average integrated gamma power per cell in naïve and experienced animals. C. Linear correlation between Vm gamma powers and spike rates for every trial (left) and for every cell (right) in the naïve and experienced groups.