1Department of Physiology, Medicum, University of Helsinki, Helsinki, Finland 2Department of Neurosciences, Clinicum, University of Helsinki, Helsinki, Finland
*Email:simo.vanni@helsinki.fi
Introduction
We have been building a phenomenological macaque retina simulator with the aim of providing biologically plausible spike trains for downstream visual cortex simulations. Containing a wide array of biologically relevant information is the key to having an accurate starting point for building the next step in the visual processing cascade.The primate retina dissects visual scenes into three major high-resolution retinocortical streams. The most numerous retinal ganglion cell (RGC) types, midget and parasol cells, are further divided into ON and OFF subtypes. These four RGC populations have well-known anatomical and physiological asymmetries, which are reflected in the spike trains received by downstream circuits. Computational models of the visual cortex, however, rarely take these asymmetries into account.
Methods
We collected published data on ganglion cell densities[1]and dendritic diameters[2, 3]as a function of eccentricity for parasol and midget ON & OFF types. Spatial receptive fields were modelled as a elliptical difference-of-Gaussians model or a spatially detailed variational autoencoder model, based on spatiotemporal receptive field data[4, 5]. The three temporal receptive field models include linear temporal filter, dynamic contrast gain control[6–8]and a subunit model accounting for both center subunit[9]and surround[10]nonlinearity and fast cone adaptation[11]. Finally, we included cone noise to all three temporal models, quantified by[12], to account for correlated background firing in distinct ganglion cell types[13]. Results
Figure 1 A and B show how synthetic receptive fields are arranged into a two-dimensional array. The temporal impulse response (C) for the dynamic gain control model has kernel dynamics varying with contrast. Parasol and midget unit responses for temporal frequency and contrast show expected behavior, with parasol sensitivity peaking at a higher temporal frequency and showing compressive non-linearity with increasing contrast (D, F). Dynamical temporal model responses to full-field luminance onset show expected onset and offset dynamics (F, G). The drifting sinusoidal grating at 4Hz evokes oscillatory response at the stimulus frequency (H).
Discussion
Our retina model can be adjusted for varying cone noise and unit gain (firing rate) levels and allows mp4 videos as stimulus input. The software is programmed in Python and supports GPU acceleration. Moreover, we have strived for modular code design to support future development. Our model has multiple limitations. It is monocular and accounts for temporal hemifield only. It assumes stable luminance adaptation state and does not consider chromatic input or eye movements. Optical aberration is implemented with fixed spatial filter. Despite these limitations, we believe it provides a physiologically meaningful basis for simulations of the primate visual cascade.
Figure 1. Fig 1. A) Synthetic parasol ON receptive fields (RFs). B) RF repulsion equalizes coverage. C) Linear fixed and nonlinear contrast gain control model temporal impulse responses. D, F) Parasol and midget unit responses for temporal frequency and contrast. E) Responses for varying contrasts. G) Responses for luminance onset and offset. H) Responses for drifting sinusoidal grating. Acknowledgements
We thank Petri Ala-Laurila for insightful comment on model construction. This work was supported by Academy of Finland grant N:o 361816.
