Data analysis

We detected saccades and microsaccades using our previously described toolbox⁸¹, and we inspected the detection results manually. To investigate whether microsaccades at image onset might have influenced the SC responses to the stimuli, whether by peri-microsaccadic modulation^45,50 or jittering of images⁴⁷, we computed microsaccade rate across time from image onset (e.g. Figure 5A,E in “Results”). We did so similarly to how we estimated microsaccade rate recently⁸. Briefly, we binned microsaccades using a 40 ms moving time window, with time steps of 10 ms. In general, we included all trials in our neural data analyses, even when there were microsaccades. This was fine because of the low likelihood of microsaccades, especially in the critical early visual burst interval. However, we also confirmed that our results were unchanged by repeating the analyses after removing all trials in which there was a microsaccade between – 100 and + 300 ms relative to image onset (e.g. Figure 5 in “Results”).

For neural analyses, we sorted the neurons offline using the Kilosort Toolbox⁸², followed by manual curation using the phy software. We then proceeded to analyze the spike rasters and firing rates.

To investigate whether SC visual responses differentiate between object and non-object stimuli, we plotted spike rasters and firing rates across the different image conditions (e.g. Fig. 3 in “Results”). We then assessed whether an ideal observer could discriminate between object and non-object stimuli just based on the SC firing rates. To do so, we performed receiver operating characteristic (ROC) analyses using 40 ms time bins moving in steps of 10 ms. In each 40 ms time bin around the time of image onset, we collected firing rates within this interval from all trials of the real-life object condition and all trials of an image control from the same neuron (e.g. phase-scrambled or grid-scrambled images). We then ran the ROC analysis to obtain an area under ROC curve measure (AUC), allowing us to assess the discriminability between the two firing rate distributions. An area under the ROC curve value of 0.5 would indicate non-discriminable firing rate distributions. We performed the ROC analyses at all times from -100 ms to + 300 ms from image onset, with 10 ms resolution. We did this because the earliest time at which the saccade target could appear in the task was 300 ms (e.g. Fig. 1). We assessed a neuron as detecting objects if its area under the ROC curve in any interval between 0 and 300 ms was statistically significantly different from 0.5. We assessed significance by calculating bootstrapped confidence intervals for the area under the ROC curve measure and using a p < 0.05 criterion. This is similar to our previous approaches³¹. Importantly, we plotted full time courses of AUC values to demonstrate the time-dependent nature of SC neural responses after stimulus onset. We then averaged across all significant neurons’ AUC time courses and obtained 95% confidence intervals across the population. We graphically labeled the time of object detection in figures as the time at which the population AUC discrimination time course first deviated significantly from 0.5 (i.e. no overlap between the 95% confidence interval and 0.5). Of course, this is not meant to be an absolute time of discrimination onset, because it is smoothed by our binning procedure. However, it does still provide an indication of whether object detection is possible in SC neurons within the time frame of their initial visual bursts or later (when feedback from other areas might emerge in the SC neural discharge); this is consistent with the example raw firing rates that we included.

In the ROC analyses described above, we pooled object categories together. Because we ran only one exemplar from each object category in a given session, it was not easy to convincingly assess whether SC neurons also exhibit early object recognition capabilities, besides detecting extrafoveal objects. Future experiments could investigate this possibility in more detail, as in, for example, the studies investigating SC face preference^15–19.

We also repeated the ROC analyses for the different functionally-classified neurons. For example, we picked only visual-motor-prelude neurons and calculated the area under the ROC curve metrics for those, or we only considered visual-delay neurons. This allowed us to assess whether early visual object detection by the SC (e.g. in the initial visual burst interval; see “Results”) only occurred in purely sensory neurons, or whether it also appeared in deeper visual-motor neurons. In some analyses, we found that whether a neuron had delay-period activity or not (e.g. visual-delay and visual-motor-prelude neurons both had delay-period activity) influenced the ROC results in either early or late intervals after image onset. Therefore, to demonstrate this point, we combined neuron types appropriately; that is, visual-motor-prelude and visual-delay neurons were combined since they both showed delay-period activity, and visual-motor or visual neurons were combined together because they both lacked delay-period activity.

In all figures and analyses, we showed results for each monkey individually.