Results and Discussion
Here we report the visual response of a population of speed responsive RGC, obtained with a 252-multielectrode array system, recorded from retinal patches from the diurnal rodent ’degu’ (Octodon degus , see Methods and Materials section). From a complete set of recorded cells, we first applied a culling based on the light-flashes and checkerboard activity (see \FIGreportA). The quality of the estimated linear response STA (amplitude \(>5\sigma\) and a good fit to equation \ref{eq:v_fit}) plus a coherent response to a brief flash () served as criteria to eliminate noisy cells in the recording.
We measured the RGC responses to a set of artificial stimuli (e.g. drifting gratings) with varying speed and spatial frequency (see \FIGreportB,C). For each drifting grating, the average response over trials was fitted with a skewed Gaussian. A subset of the complete set of valid RGC was separated and classified as speed responsive cells , confirmed by cells whose response to the speed of the drifting grating has a good fit to \EQv_fit at every spatial frequency tested (\(\chi^{2}<0.05\) for the normalized response) \FIGreportC. We analyzed retinal patches from two animals, where \(144\) and \(164\) speed responsive cells were found, respectively, representing \(\approx 46\%\) and \(\approx 44\%\) of the total RGC registered.
Most speed responsive cells can be classified as preferring either lower or higher speeds. In \FIGreportB and supplements, we show how preference for certain speeds changes as a function of spatial frequency. As expected for RGC, preferred speed decreases at higher spatial frequencies, but the tuning range varies from cell to cell.
Speed selectivity of RGC differs when they are confronted with enriched stimulus with energy distributed along a larger area of the parameter space (Motion Clouds). The results presented so far suggest that the response properties of observed RGC in degus match with those reported for other non-primate mammals \citep{Troy:2002tq,Balasubramanian:2009jx,Tang:2016ik}. However, when presenting a stimulus with the same base properties but with energy distributed along a larger area, the responses differ significantly. For many cells, the response profile to the MC stimuli is smaller along both axes when viewing it as a two dimensional map of the response at each combination of parameters (\FIGreportC). Examining in detail the response to the preferred spatial frequency (\FIGreportD) shows that the response to lower and higher speeds is much weaker for the MC stimuli regarding the response to the grating, while the responses become equivalent for firing rate only around the preferred, intermediate speeds. As mentioned, natural stimuli elicits stronger but sparser responses in the visual cortex \citep{Vinje:2000fk} and narrower tuning curves for stimulus containing complex mixtures of signals \citep{Priebe:2006jk}. How these two phenomena are linked remains unsolved, but it has been related to the tuning of the visual system to natural scenes and efficient coding \citep{OLSHAUSEN:2004fw}.
To quantify the change in the tuning bandwidth of the grating versus MC response, we measured the properties of the curve fitted to the response to speed (see Methods section for details). In \FIGdifftuningA we show some examples of RGC with narrower tuning in terms of lower response to the non-preferred speed. At the population level, as can be seen in \FIGdifftuningB, C and D, a large proportion of the speed responsive cells show a decrease in the bandwidth measured as \(\sigma_{v}\) for the complex stimuli when compared to the response to drifting gratings. This is evidenced when looking at the comparison of \(\sigma_{v}\); for both MC series a large portion of points falls below the line of equality, and when looking at the distribution of the magnitude of differences (\(\Delta\sigma_{v}\)), the distribution is skewed toward positive values. To quantify the differences in the distribution, we separated the central bin (unchanged bandwidth) from the cells with increased and decreased bandwidths (\FIGdifftuningD). As can be seen, cells with narrow tunings for the MC represent a significant proportion of the total speed responsive cells . However, in contrast to what is observed in the cortex, cells responding to a narrower range of parameters do not present a higher firing rate compared to grating response, except in rare cases (3 cells among the 144 responsive to speed).
Such narrower tuning can have at least three implications. First, models based on the response to simple stimuli fail to adjust to the response to complex stimuli, meaning that it is not possible to directly extrapolate response properties. In turn, this proves once more that even hough significant research on the early visual system exists, we still do not grasp the full extent of its functioning \citep{Carandini:2005kw,Masland:2007wi}.
The second implication relates to the possible mechanism behind the change in the tuning to speed. In the present work, the difference between each set of stimuli (and thus the factor that would determine the differences in the response) is the spatial frequency content, i.e. the variety in size of the elements that constitute the images; while the three classes of stimuli have the same average properties (spatial frequency and speed) for each set, the MC contain additional signals or components around this central value. It has been shown that stimuli that completely fail to modify the activity of a cell when presented in isolation, can modify the response to another stimulus when presented simultaneously \citep{Vilankar:2017jl}. This would stem from nonlinearities in the processing of the signal performed by the cell, in the sense that the response to a complex stimulus is not simply the response to the sum of its components. This way, the additional signals contained in the complex stimulus would act as inhibitor, resulting in lower response when the parameters are farther from the preference of the cell \citep{Carandini:2011fm} and more reliable code \citep{Cafaro:2010im} (less variability between trials). Interestingly, together with this narrower bandwidth, many cells show an increase in preferred speed (see two examples in \FIGdifftuningA, left panels), while others preserve their preference (\FIGdifftuningA, right panels). The change in the speed preference cannot be attributed to changes in the spatial frequency content of the stimulus. Due to the bandwidth of the stimuli, the images will include additional spatial frequencies, but the shape of the spectrum (see \BOXmotionclouds) determines that most of the additional frequencies will be higher than the central frequency (\(\text{sf}_{0}\)). If the cell processed these higher spatial frequencies linearly, they should shift preference towards lower speeds - but our results show the opposite. From our analysis, it is difficult to relate or predict this behavior from other properties of the cell response, since the shift in the preferred speed does seemingly not correlate with any receptive field characteristics (\FIGreport and supplements). Many types of nonlinearities explain the response to naturalistic stimuli \citep{Wegmann:1999bs}, but regardless of the specific mechanism, this change in response can be explained with hyper-selectivity \citep{Golden:2016hh,Vilankar:2017jl}. Hyper-selectivity is a property described for visual neurons in the cortex. It explains the effect of many types of nonlinearities in the response to stimuli into a single framework based on a geometric description of the cell’s Receptive Field (RF). In this paper, we show that such a mechanism would appear as early as in the retina, highlighting another example of the complex functions this tissue performs \citep{Gollisch:2010kv}.
A third effect relates to the consequences of the functional origin of these narrower tunings for naturalistic stimuli. At the individual cell level, it means the cell will only respond when the stimulus occurs near the preferred parameters. But more importantly, when looking at the population code, narrower tuning curves can lead to less overlap between cells, so in conjunction the code will contain less redundancy, an important aspect of the efficient coding theory \citep{Barlow:1961ww,Barlow:2001ub}. As such, this results mean that for speed-selective cells, an input with a larger spectrum of spatial frequencies will have a higher precision in speed, independently to the local scale (spatial frequency) of the stimulus. This mechanism speaks for the dual mechanisms at the origin of selective responses (here speed) and of the invariance to other features (here scale). However, it is still unknown how these features emerge in the low-level visual system, though a computational study predict that this could be explained using unsupervised learning models \citep{OLSHAUSEN:2004fw}.
In conclusion, we have proved that the retina would be adapted to at least some of the characteristics of natural images, specifically those related to motion and spatiotemporal information content and correlations. This adaptability manifests in the form of narrower shapes of the tuning curves, which leads to less overlap between cells in the feature space and thus reduced redundancy in the population code and lesser firing rate, which translates into lower energy expenditure and less channel saturation.
Methods and Materials
\(\)
Electrophysiological recordings
Octodon degus were raised in captivity in a controlled facility. Experimental procedures are approved by bioethics committee and regulations from the University and in accordance with the bioethics regulation of the Chilean Research Council (CONICYT). Degus are diurnal rodents with a 30% of their photoreceptors being cones. For a complete description of the model see \citep{Chavez:2003bn,Jacobs:2003dd}. The physiological response of RGC from the Degu was measured using different types of visual stimuli troug\citep{Litke:2004bc,Segev:2004cu} (USB MEA256, Multichannel Systems) and the experimental protocol follows \citep{PalaciosMunoz:2014gi}. Briefly, prior to an experiment animals were put in darkness for \(30\min\), then deeply anesthetized with halothane and beheaded. Eyes were removed at room temperature and the cornea removed under red illumination. Thereafter the posterior hemisphere of the eye was dissected in quadrants and the pigmented epithelium separated from the retina. Finally, a piece of retina was set upon a dialysis membrane (Spectra/Por Dialysis membrane 132554 MWCO: 25,000, Spectrumlabs) and mounted on a perfusion chamber, which was then lowered onto the electrode array with the RGC side down. Recording commenced under perfusion with AMES medium bubbled with 95% O2 5% CO2 at 33º and the pH adjusted to 7.4.