P107 Towards A New Metric for Understanding Neural Population Encoding in Early Vision
Silviya Hasana*1, Simo Vanni1
1Department of Physiology, Medicum, University of Helsinki, Helsinki, Finland
*Email: silviya.hasana@helsinki.fi
Introduction Representational fidelityofvisioncan be evaluatedusing adecoding approach[1-3],howeverthe methodisinexplicableand difficult to quantify.This study aims to develop a quantitative metric for vision models based on neural population spike activity.By analyzing the relationship between stimulus features, spike timing, and receptive field locations, we investigate how population spike data encode information.We apply both deterministic and probabilistic approaches to evaluate neural populationsencoding,andsubsequentlyquantitativedecodingcapacityonspatialstimulus information.Aquantitativeperformancemetricis fundamental for advancing functional computational vision models, such as the SMART system[4].
Methods We useda macaque retina model with simulatedON and OFFparasolunitsin a 2D retinal patch.Stimuli were varied for spatial parameters, such asspatial frequency and orientation.First, we summed the spikes generated within 500msof grating stimulus onset for each unit separately.Then, wecalculated the difference between ON- and OFF-unit spike counts and normalized the responses.Toevaluatewhether activation patterns aligned with thetruestimulus, we binned the responses and applied a deterministic Gabor filteratdifferent orientations.Subsequently, we plan to evaluatemodel performanceusing a Bayesian Ideal Observer, which models prior, likelihood, and posterior as a tuning curve foroptimalstimulus decoding.
Results Our findings showed the presence of orientation-specific patterns in neural population activity, both in the deterministic and probabilistic approaches. Based on preliminary data analysis and processing in our deterministic approach, we expected a strong match between Gabor kernel prediction and true orientations for parasol ON and OFF cells. Our experimentstested100 sweeps, obtained 100% accuracy for oblique orientation prediction with mean average errorof2.97degrees. The high accuracy of the deterministic approach confirms that simple feature-based encoding mechanisms, such as Gaborfiltermatching, align well with neural responses in the parasol ON and OFF cells.
Discussion As expected, the modeled retinal ganglion cell population encodes orientation in a structured manner that can be decoded based on receptive field positions in the visual field.Moving forward, we will explorea probabilistic approach by applying curve tuning through a Bayesian Ideal Observer to assessthe reliability of neuron population spike activityencodes stimulus orientation, spatial frequency, and motion direction.The probabilistic approach will incorporate prior and likelihood to reconstruct stimulus orientation. The resultswillassesshowboth deterministic andprobabilisticapproaches complement, andcontributetoneuraldecoding, providing a quantitative metric for evaluating functional vision models.
AcknowledgementsThis work has been supported by Academy of Finland grant No: 361816 References [1]https://doi.org/10.1371/journal.pcbi.1006897 [2]https://doi.org/10.1038/s41583-021-00502-3 [3]https://doi.org/10.1038/nrn2578 [4]https://doi.org/10.1016/j.brainres.2008.04.024
