First spikes in visual cortex enable perceptual discrimination

Visually guided perceptual decisions involve the sequential activation of a hierarchy of cortical areas. It has been hypothesized that a brief time window of activity in each area is sufficient to enable the decision but direct measurements of this time window are lacking. To address this question, we develop a visual discrimination task in mice that depends on visual cortex and in which we precisely control the time window of visual cortical activity as the animal performs the task at different levels of difficulty. We show that threshold duration of activity in visual cortex enabling perceptual discrimination is between 40 and 80 milliseconds. During this time window the vast majority of neurons discriminating the stimulus fire one or no spikes and less than 16% fire more than two. This result establishes that the firing of the first visually evoked spikes in visual cortex is sufficient to enable a perceptual decision.


28
Perceptual decisions involve the sequential activation of several, hierarchically organized 29 cortical areas beginning with early sensory areas and ending with associational and motor 30 areas. Based on the number of areas likely involved in the processing of sensory stimuli it has 31 been hypothesized that in each area a relatively brief time window of activity may be sufficient 32 to enable a perceptual decision (Fabre-Thorpe, Richard & Thorpe, 1998). Yet, this time window 33 has never been directly measured for any specific area. By determining these lower limits and The lack of answers to these questions is largely due to technical limitations. One key issue is to 43 demonstrate that the visual area of interest is necessary for the sensory discrimination task at 44 hand. Even though activity in a given area may carry relevant stimulus information, that area 45 may not be required for the perceptual decision. A second challenge is to precisely control the 46 duration of the sensory evoked response of that visual area. Answering this question has been 47 technically difficult since the duration of visually evoked activity in the brain cannot be precisely 48 controlled by the duration of the sensory stimulus. Even a stimulus as brief as 16 ms triggers a 49 response that lasts hundreds of milliseconds in visual cortex (Rolls, Tovee & Panzeri, 1999). 50 the side of stimulus presentation (e.g. Figure 2A) and allowed the animals to recover for ten days 154 post-surgery before behavioral testing. Lesioned animals performed at chance (p>0.3, Wilcoxon 155 rank sum test on choice data, n=4 mice, Figure 2B). The impairment in behavioral performance 156 was not due to the ten day interval from the last behavioral session because trained control 157 animals experiencing even longer intervals between behavioral sessions remained proficient 158 (accuracy of 80±10%, Figure 2C). Furthermore the behavioral impairment was not due to either 159 anesthesia or to some unspecific impact of surgery because proficiency was preserved after 160 removing the ipsilateral VC (accuracy of 85%, n=1 mouse, Figure 2B) or following anesthesia to 161 perform craniotomy for physiological recordings (80±10%, see results below). Taken together, 162 these results show that this visual discrimination task requires visual cortex. 163 Neurons in primary visual cortex report stimulus identity by 80 ms 164 To determine over what time interval stimulus evoked spiking activity in individual V1 neurons 165 can be used to disambiguate the target from the distractor stimulus we recorded extracellular 166 action potentials while the animals performed the task ( Figure 3A). We inserted a multichannel 167 probe in V1 at the beginning of a behavioral session in trained mice (performance accuracy 168 during recordings: 80±10%, mean ± std; n = 9 mice). To ensure that the units were maximally 169 excited by the stimulus, we placed the monitor so that the position of the stimulus in the initial 170 350 ms, when the stimulus is stationary, was superimposed on the multiunit spatial receptive 171 field (center of stimulus was 2±1 degrees from center of receptive field, mean ± std, n=8 mice). 172 The cortical response to the visual stimulus began 40±5 ms after stimulus onset (mean ± std 173 across mice, Figure 3C) consistent with previous reports (Niell & Stryker, 2008). The onset of 174 cortical response was quantified as the earliest deflection in the local field potential that 175 exceeded 3 standard deviations from baseline. We verified that the earliest deflection 176 corresponded to layer 4 of V1, the major thalamo-recipient layer, based on current source density 177 analysis (Niell & Stryker, 2008) (Supplementary Figure 3B). 178 To determine whether the spiking of an individual neuron allows an ideal observer to 179 discriminate the target from the distractor stimulus we performed ROC analysis (Tolhurst, of the cortical response ~50% of discriminating units discriminate the target from the distractor. 196 To determine how well the orientation tuning curve of a neuron predicts its ability to 197 discriminate we measured the tuning properties of discriminating neurons after the end of the 198 behavioral session. We presented drifting gratings of twelve different orientations that had the 199 same size and spatial frequency and were presented at the same location as the stimuli used discriminating units that preferred the target during the task showed a peak response to 204 orientations larger than 90 degrees (109 ± 8 degrees, median ± SEM; n= 9, 4 mice; Figure 3G). 205 Furthermore most discriminating units that preferred the distractor during the task, showed a 206 peak response to orientations less than 45 degrees (30 ± 20 degrees, median ± SEM; Figure 3G;  Figure 3H).
The threshold duration of V1 activity for perceptual discrimination limits most neurons' 216 firing to one or no spikes. 217 What is the minimal duration of activity in visual cortex necessary for accurate visual 218 discrimination? And how many action potentials are fired by individual neurons during this 219 time? If by 80 ms from the onset of visually evoked cortical activity information about stimulus 220 identity is available to an independent observer, it may also be available to the mouse. Thus, the 221 minimal duration of visual cortical activity enabling discrimination may be around 80 ms.

222
To control the duration of the visually evoked cortical response we optogenetically silenced 223 visual cortex, as described above, at varying intervals after the onset of the response ( Figure 4A).

224
In each experiment we ensured that the LED intensity was sufficiently high such that 225 performance accuracy was at chance when the illumination started before the stimulus appeared 226 (p>0.05, Wilcoxon ranksum test on choice data, n=8 mice). Furthermore, as above, for each 227 animal we verified that despite chance performance the hold times of the stimulus in the reward 228 zone of the monitor did not differ between target and distractor stimulus (p>0.05, Wilcoxon 229 ranksum test on stimulus centering times in the reward zone). We verified this again at the very 230 end after testing all LED onset intervals.

231
The accuracy of the behavior increased with increasing interval between the onset of the cortical 232 response to the stimulus and the onset of cortical silencing ( Figure 4B,C). When cortical 233 silencing followed the onset of cortical response by 44±6 ms the performance was close to 234 chance (54±5%; mean ± std across mice, Figure 4D), similar to when the LED onset preceded 235 stimulus presentation (51±3%; mean ± std across mice). Strikingly, however, when cortical 236 silencing was delayed by a further 40 ms, hence with a latency of 80 ms after the onset of the 237 cortical response, performance accuracy of the animals sharply increased to 76±7% (mean ± std 238 across mice). Performance accuracy continued to increase, yet less sharply, over the longer 239 intervals tested reaching 92±5% when the LED onset followed the onset of the cortical response 240 by 300 ms (mean ± std across mice). With this interval the animals performed similarly to 241 control conditions, in the absence of LED illumination (94±2%; mean ± std across mice). Thus, 242 there is a sharp increase in performance when visual cortex is allowed to function between 44 243 and 80 ms after the onset of the cortical response. As above, we used ROC analysis to compare 244 behavioral performance with the ability of an ideal observer to disambiguate the target from the 245 distractor based on times spent by each stimulus in the reward zone when silencing cortex at 44 ms following the onset of the cortical response. The discrimination accuracy of the ideal 247 observer was 54±7%, hence very close to the actual performance of the task at 44 ms (54±5%). 248 These results show that the minimal duration of visually evoked activity in V1 for an animal to 249 perform the present task above chance lies between 40 and 80 ms.

250
If the estimated time window indeed approximates the threshold duration of V1 activity for 251 perceptual discrimination, performance accuracy in trials when V1 is active for only 80 ms 252 should be very sensitive to the difficulty of the task. We thus trained mice to discriminate a 253 narrower angle difference between target and distractor, namely 15 degrees. Mice were first 254 trained to perform the standard 45 degrees discrimination task and their behavioral performance 255 measured across various intervals of cortical silencing, as above. We then re-trained those same 256 animals to discriminate a target from a distractor separated by 15 degrees until they reached a 257 similar level of proficiency as for the 45 degrees task (accuracy of 90±4% for 15 degrees versus 258 accuracy of 93±2% for 45 degrees, mean±std, n=3 mice, Figure 5B). We silenced the cortex of 259 these animals at various intervals following the onset of the cortical response and compared the 260 decrease in performance between the 45 and the 15 degrees discrimination tasks. Silencing 261 cortex at 80 ms after the onset of the cortical response reduced performance significantly more 262 for the 15 degrees as compared to the 45 degrees discrimination task in all animals (p<0.05,

263
Wilcoxon ranksum test on choice data, n=3 mice Figure 5B,C). While silencing V1 80 ms 264 following the onset of the cortical response still enabled the 15 degrees discrimination to occur 265 above chance (p<0.02, Wilcoxon ranksum test on choice data, n=3 mice), the accuracy was 266 significantly lower than for 45 degrees discrimination (p<0.02, Wilcoxon ranksum test on choice 267 data, n=3 mice, Figure 5B,C). This difference cannot be accounted for simply by a difference in 268 motivation or in control performance because in two out of three mice non-LED trials during the 269 15 degrees discrimination task were as accurate as non-LED trials during the 45 degrees 270 discrimination task (Wilcoxon ranksum test on choice data, p=0.65 and 0.67, Figure 5B). Thus, 271 these experiments demonstrate that the time between 40-80 ms following the onset of the cortical 272 response indeed captures the threshold duration of V1 activity for a simple perceptual 273 discrimination.

274
Over these initial 80 ms from the onset of the cortical response discriminating units in primary  Figure 3C), corresponding to a firing rate of 7.5 Hz, and 0.22±0.06 action 280 potentials for their non-preferred stimulus. Furthermore, in response to their preferred stimulus, 281 discriminating units fired only 1 or no action potential in 80% of the trials and 2 action potentials 282 in only 11% of the trials ( Figure 6E), similar to what is expected by Poisson statistics (% of 283 variance explained across units: R 2 =97±5%, median ±SEM; median time until first action 284 potential: 70±10 ms and 80±20 ms from the onset of the cortical response for the preferred and 285 non-preferred stimulus, respectively (median ±SEM across units; analysis performed over the 286 initial 300 ms from the onset of the cortical response, Figure 6C); mean latency difference: 12± 6 287 ms (mean ±sem across units; p=0.03; t-test; Figure 6D)). Thus, over the initial 80 ms from the 288 onset of the cortical response the vast majority of discriminating units in primary visual cortex 289 get to fire either one or no action potentials.

290
To determine whether indeed the first action potential in response to a stimulus is sufficient to 291 discriminate the target from the distractor we performed ROC analysis ( Figure 6G) after 292 removing from each unit all but the first action potential after the onset of the cortical response.

293
As above we performed this analysis for various intervals from the onset of the cortical response.

294
The first action potential was sufficient for ~33% of units to discriminate by 300 ms (compared 295 to 46% if all the action potentials were available), and more than half of those units (54%) could 296 discriminate above chance at 80 ms ( Figure 6H). Thus for most units the first action potential 297 substantially contributes to their ability to discriminate within the initial 80 ms after the onset of 298 the cortical response.

299
Finally, the accuracy of the behavioral response during the initial 80 ms can be explained by

308
We have developed a visual discrimination task that necessitates visual cortex because both 309 acute cortical silencing and permanent ablation reduces performance of the task to chance. By   contribute to the discriminability of the stimulus (compare Figure 6G with Figure 3E).

333
The ability for a neuron to disambiguate two stimuli with only one or no spike depends on how 334 distinct the response of that neuron is for those stimuli and on the trial to trial variability of its    We have provided direct evidence for the minimal amount of time that it takes visual cortex to 416 process visual information in order to enable a perceptual decision and determined the neuronal 417 activity that occurs during that period. The speed at which humans are able to discriminate visual 418 stimuli has led to the suggestion that processing of the visual stimuli can be accomplished with 419 individual neurons in each of the relevant brain areas firing either none or one action potential. 420 This work demonstrates that a period of activity in mouse primary visual cortex during which 421 most neurons fire none or one action potential is indeed sufficient to enable perceptual 422 discrimination. Future work will elucidate which downstream brain areas read out these first 423 essential spikes generated in V1.  2007)) was marked at the surface of the skull. Using a dental drill (700 µm) the area of visual 455 cortex was thinned and removed. Sterile PBS was used to hydrate the exposed brain area. A cut 456 of 1 mm depth was performed around the outline of VC using a microsurgical blade (FST). The 457 cortical tissue was removed using a spoon shaped microsurgical blade (FST 10317-14). The area 458 was washed with PBS to remove blood and consequently covered with Silicon Kwik-Cast (WPI).

459
Upon polymerization a layer of cyanoacrylate glue was applied to cover the lesioned area. An 460 additional layer of dental cement was applied to permanently cover the lesioned site.    Table   492 1). The distance between two consecutive stimuli in the track was 1.25 times the width of the 493 monitor.
For the recording sessions, the spatial phase of the grating was constant and did not vary trial to   Behavioral training. Training began after animals had been on water restriction for at least 7 523 days (~1 ml water/day). During training, mice were kept at 80% or above of their initial body 524 weight. Additional water was provided if the body weight fell below 80% of the initial weight.
The initial behavioral parameters were: 100% contrast, gain (gain= stimulus displacement in the 526 monitor (cm)/ running distance (cm)): 0.6, hold time (minimal time in reward zone for a reward): 527 0.2s. With these parameters, mice would get a water reward every time a new target stimulus 528 would pass the reward zone, i.e. as long as they kept running. After mice began to run 529 consistently (one to a few sessions), the gain was decreased to 0.45 and the hold time was 530 initially increased to 0.4 sec to get the first ~10 rewards each session (i.e. during the warm up 531 period) and then increased to 0.9 sec. Mice learned to perform the task, that is to hold the target 532 in the reward zone for at least the minimal hold time for a reward, but not the distractor, with when the LED illumination started before the stimulus appeared, the animal's discrimination 547 accuracy was similar to that of an ideal observer based on ROC analysis of the stimulus 548 centering times in the reward zone. However, when the LED illumination started after the 549 stimulus appeared, particularly for intervals longer than 80 ms from onset of the cortical 550 response, the animal's discrimination accuracy was usually noticeably lower than that of an ideal 551 observer (>10% difference). This difference would often occur because on some target trials 552 mice would not slow down sufficiently for the trial to be a 'stop trial' but they would slow down 553 more than they would for distractor trials. Thus, an advantage of our task is that it revealed 554 differences in the animal's behavior for target versus distractor that were not captured by the 555 binary classification of stop versus non-stop trials.
To motivate mice to make choices with similar accuracy to that of an ideal observer monitoring 557 the time that the stimulus spends in the reward zone, we adjusted the probability that an image 558 would be a target and the minimum time that the target had to be centered for a reward (hold  The probability that a stimulus would be a target varied from 25%-50% across mice, and the 571 hold time varied from 0.6s -1.0s across mice. Each interval was tested for 1-3 sessions totaling 572 130±70 trials (range: 42-372 trials per interval), and data were pooled together for analysis.  Data collection began at least 30 min after insertion of the probe. We first presented black or 585 white squares of ~10º to map the location of the receptive field across all channels of the probe.

586
To ensure that all receptive fields overlapped with the stimulus position during the behavior in the first 350 ms of a trial ( Figure 1A), we either moved the monitor slightly if the movement was 588 approximately <15º, or reinserted the probe at a different location mediolaterally. We mapped 589 the receptive field at a higher spatial resolution at the end of the recording for 8 mice (for same 590 units and same stimulus and monitor position, see 'Visual Stimulation').   (Figures 1 and 3). A stop trial is 602 defined as the stimulus spending ≥ the minimal time for reward in the reward zone (500 pixels 603 wide). Error bars in stop probability (Figures 1, 2, 4 and 5) indicate 95% confidence intervals 604 assuming a binomial distribution of stops and non-stops at each orientation (data pooled from all 605 sessions that each condition was tested). Accuracy for the target stimulus is defined as the 606 percentage of stop trials upon target presentation; accuracy for the distractor stimulus is defined 607 as the percentage of non-stop trials upon distractor presentation. Overall accuracy is taken as the 608 average of these two choice accuracies; chance level is 50% correct. field to the center of the stimulus was 2±1 degrees (mean ± std, n=8 mice). possibility that this analysis was not sensitive enough for low firing units, out of 98 well isolated 650 units, 13 (all regular spiking) were excluded because they fired < 1 spike every 6 trials over the 651 initial 300 ms. This threshold was chosen because units firing at rates just above this threshold 652 could discriminate (p<0.012; Wilcoxon ranksum test comparing the distributions of the number 653 of action potentials for target versus distractor). We confirmed that the distribution of running 654 speeds for the two stimuli was not significantly different in the initial 350 ms and thus did not 655 affect our ROC analysis (p>0.02, Wilcoxon ranksum test using the Benjamini-Hochberg 656 correction for multiple comparisons, n= 9 mice).

657
To determine how many neurons are needed to explain behavior, we first artificially increased  Figure 6C). We also show the mean of the spike times for 672 each unit and the distribution of these means across units ( Figure 6D). To assess whether the first 673 spike occurred earlier for the preferred versus the non-preferred stimulus, for each unit we 674 computed the difference in the mean time of the first spike for the preferred versus the non-675 preferred stimulus and tested whether the mean of the differences from all units was different 676 than zero (Student's t-test).

677
To compute orientation tuning curves, the firing rate for each orientation was calculated over the 678 initial 330 ms following cortical onset (i.e. the first cycle of presentation), averaged across 679 repetitions, and normalized by the maximal firing rate across orientations.

680
To compute the preferred orientation for each unit, we used the following equation (Lien &