Note that mice were free to run during playback sessions and even showed a tendency to match running to playback of visual flow—that is, in some experiments, onset of visual flow induced the animal to run. In addition, in the closed-loop configuration, visual feedback was perturbed by briefly halting visual flow for 1 s at random times Regorafenib manufacturer (Poisson distribution, with a probability of 0.25% of perturbation every 30 ms time bin, with a refractory period of 3 s). Animals

were briefly (approximately 10 s) anesthetized with isoflurane for head fixation on the ball. Animals were then allowed to get used to the head fixation and the setup and were exposed to normal visual feedback for 10–30 min. Experiments consisted of a 2 min feedback session with 1 s perturbations of visual flow at random times and a series of three to four playback sessions, each typically consisting of 2 min feedback without perturbations and 2 min

of replay of the visual flow generated by the mouse during the first 2 min feedback session (including perturbations). In initial experiments, data were not recorded during the 2 min feedback without perturbations between playback sessions. Thus, each experiment consisted of 8–16 min of recording. During all experiments, the animal’s left eye was filmed with a video camera (The Imaging Source, ATM/ATR tumor 30 fps). Pupil position and pupil diameter were extracted online with custom-programmed software based on the design of Sakatani and Isa (2007). Two-photon images were full-frame registered using a

custom-written registration algorithm. The standard deviation of brain displacement parallel to the imaging plane was 2.4 ± 0.3 μm (mean ± SEM). All data in which cells visibly moved perpendicular to the imaging plane were discarded. Cells were selected based on mean and maximum projections of the data by hand (typically the nucleus was excluded from the selection). Use of the maximum projection ensured the inclusion of all active cells, even ones that were not visible in the mean projection. This slightly biased our cell selection toward active cells. Fluorescence L-NAME HCl traces were calculated as average fluorescence of pixels lying within the cell in each frame. To remove slow signal changes in raw fluorescence traces, we subtracted the 8th percentile value of the fluorescence distribution in a ±15 s window from the raw fluorescence signal (Dombeck et al., 2007). In addition, signals were low-pass filtered at 10 Hz. ΔF/F signals were calculated by dividing raw fluorescence signal by the median calculated over the entire fluorescence distribution of each cell and then subtracting 1 from this value. To minimize the influence of movement artifacts on average fluorescence measurements, we estimated the movement-related signal noise by calculating the standard deviation σMN of the lower half of the fluorescence distribution (ΔF/F < median[ΔF/F]) for each cell individually.