Decoding of complex movies from a large retinal population Vicente Botella-Soler 1 *, St´ ephane Deny 2 , Olivier Marre 2 , Gaˇ sper Tkaˇ cik 1 1 Institute of Science and Technology Austria, Klosterneuburg, Austria 2 Institut de la Vision, INSERM UMRS 968, UPMC UM 80, CNRS UMR 7210, Paris, France * Presenting author: vbsoler@ist.ac.at Decoding of complex stimuli from the retinal activity remains an open challenge. To date, experiments have focused on decoding either a small number of discrete stimuli (e.g., decoding among several possible orientations of a drifting grating), or very low dimensional dynamical traces (e.g., of luminance in a full field flicker experiment). Here we stimulated a rat retina with a new class of synthetic stimuli, in which between 2 and 10 dark circular spots executed random dynamical motion on a square domain with no other constraints except for hard-core repulsion between the spots and reflection from the domain boundaries. The radius of the spots was ∼ 100μm on the retina, somewhat smaller than the typical center of the ganglion cell receptive fields; to achieve good performance, the decoder would thus need to make use of the population code. We constructed separate decoders densely tiling all spatial locations in the stimulus. We then reconstructed the entire movie by decoding separately these local luminance traces from the spiking activity of ∼ 100 neurons simultaneously recorded from a dense retinal patch. As a baseline, we trained linear decoders, whose performance was impacted by the substantial amounts of spontaneous activity in the rat retina. To mitigate this problem, we constructed nonlinear extensions to these decoders, which reached cross-validated correlation of > 80% in locations with good spatial coverage in the recorded population. This is achievable using biologically realistic sparse decoding kernels for each location; the resulting “decoding fields” of retinal ganglion cells are sharply localized, exhibit fine structure on a ∼ 50μm scale, are collocated with the corresponding receptive field centers, and generalize across stimuli with different numbers of randomly moving spots. Decoding paradigms with rich dynamical stimuli can thus complement encoding studies to provide novel and practical insights into the organization of the neural code. Additional detail 200 µm Figure 1: A zoom-in of the real (left) and reconstructed (right) single frame of the stimulus movie. Four randomly moving spots present in the stimulus are clearly distinguishable in the reconstructed frame. For reference, the average 1-std center contour of the ganglion cell receptive field is shown in the left panel (dashed line).