Details of the recordings and stimulation can be found in Supplemental Experimental Procedures. Data acquisition was controlled with custom-made software, written in Visual C++. Incoming data were both stored for offline analysis as well as directly processed in an online fashion. After visual inspection of the voltage signals of all available channels, one channel was selected that displayed large, homogeneous spike shapes. For this channel, an amplitude AZD5363 threshold was determined, based on a 1 min recording under stimulation with broadband flickering light intensity, to separate spikes from background

noise (Figure 2B). Only units whose spike amplitudes were well separated from the noise and that showed a clear refractory period were used for further investigation. To

verify that the simple online spike detection and sorting worked well, we occasionally performed additional offline Gefitinib in vitro analysis spike sorting, based on the detailed spike shapes (Pouzat et al., 2002). This confirmed the results obtained directly from the online analysis. To identify the spatial receptive field of a recorded ganglion cell, we first used online analysis to find the midlines of the receptive field in two orthogonal directions. Each midline was determined by dividing the stimulation area by a separation line and comparing responses from stimulation on each side of the line individually. The separation line was then iteratively adjusted until both sides yielded the same response. Finally, receptive field size was determined with blinking spots centered on the crossing point of the two identified midlines. To measure an iso-response 3-mercaptopyruvate sulfurtransferase curve, we first selected a predefined response (either average spike count or average first-spike latency). The response selection typically aimed at requiring around 30%–70% contrast for the predefined response from stimulation of one receptive field half alone. Using this range largely avoided coming too close to the physical limit of 100% contrast along the iso-response curve and

at the same time provided enough contrast for reliable spike responses. Each data point of an iso-response curve was then obtained by performing a simple line search along a radial direction in stimulus space. Details about the closed-loop experiments and search algorithms are given in Supplemental Experimental Procedures. We quantitatively analyzed the shape of the iso-response curves in two ways. To determine the degree to which curves were convex or nonconvex (Figures 3G–3I), we calculated form factors that compare the central region of the iso-response curve to the linear prediction that is obtained from the two intersection points of the curve with the axes. The form factor is larger or smaller than unity, depending on whether the iso-response curve is convex or nonconvex, respectively.