1. We have extended our simulations of neurons in 49 cortical columns to include a neuromorphic device that intermittently records from and stimulates 4 cortical columns. The device is meant to create a second, desired visual geometry in which discrete phosphenes are unified. Multinomial logistic regression analyses of spike frequency distributions from simulations that do and do not include the device classify expected numbers of phosphenes with accuracy greater than 99%. This result is important because it demonstrates that adding a second, desired visual geometry with the goal of unifying discrete phosphenes alters spike frequency distributions in a way that permits correct classification of the number of expected phosphenes.
2. We have performed additional multinomial logistic regression analyses of spike frequency distributions which classify expected perceptions of local lightness maxima (phosphenes) on one or more regions of the visual geometry (locations in visual space). The lowest classification accuracy is 98.99% correct for a control parameter value of 40 spikes/s. Accuracy is 100% correct for control parameter values of 50 spikes/s and 60 spikes/s. These results are important because it seems likely that information on local lightness maxima on regions of the visual geometry underlies classifications of the number of phosphenes.
3. We have devised three methods of visualizing spike frequency distributions in an effort to understand properties of the distributions that underlie the above classifications. One method demonstrates that distributions contain information on the column or columns which receive electrical stimulation above the phosphene threshold. This information is exceptionally clear when the control parameter = 60 spikes/s, as shown by the example labeled Figure A. The second method reveals differences between distributions that are produced by changes in synaptic strengths that are introduced in order to alter the visual geometry and thereby unify discrete phosphenes. This method is illustrated by the example labeled Figure B. The third method reveals the dominant difference or differences between distributions that are produced by altering synaptic strengths. This method is illustrated by the example labeled Figure C. All three methods are effective for network alone simulations and for simulations that include the neuromorphic device. These results are important because knowing the columns stimulated and the visual geometry is sufficient for identification of local lightness maxima on regions of the visual geometry.
Surfaces estimated from mean excitatory neuron spike frequencies for simulations of 10 seconds duration are shown for the case of stimulation of columns 5 and 12. The data are from simulations that include the neuromorphic device. The locations of columns 5 and 12 are highlighted using semi-transparent cylinders. Note that the graphic has been rotated through 180° (turned upside down) in order to reveal the low frequencies in stimulated columns. This visualization is made possible by subtracting mean spike frequencies obtained from simulations in which no columns receive above phosphene threshold stimulation. The graphic in the left panel illustrates frequencies for simulations that use the normal geometry, and the graphic in the right panel illustrates frequencies for simulations that use an altered geometry in which synaptic strengths have been changed so that the visual regions corresponding to columns 4 and 5 are each set equal to the union of regions 1 and 4 and 5.
Array plots are shown for the mean excitatory frequencies depicted in Figure A in order to make possible visualization of differences between the frequency distributions for the normal (left) and altered (right) visual geometry data. The correspondence between column numbers and positions in the arrays are given in the table in the rightmost panel of this figure. The color spectrum shown between the colored array graphics is anchored to the minimum and maximum frequencies for all stimulation conditions. Note that neurons in columns 5 and 12 have the lowest frequencies. The positions in the arrays that are labeled using the integers 1-5 are examples of visually obvious differences in frequencies that arise from differences in the visual geometries.
A visualization that captures the dominant differences between the frequency distributions for the normal and altered visual geometry data is shown. The absolute value of the ratio of the difference in spike frequencies for the normal and altered geometries that are plotted in Figure A is computed and displayed as an estimated surface. In this example the peak occurs over column 5. Not all stimulation conditions produce such large maxima in ratios that are localized at one or more stimulated columns. However, almost all conditions produce easily visualized maxima.
This patented subject matter and technology, titled “Sense Element Engagement Process of Cortical Prosthetic Vision by Neural Networks,” relates to cortical prosthetic vision processes and, specifically, to production of cortical prosthetic vision by neural networks via a sense element engagement process.
Covered by: US Patent No. 11,769,034; PCT International Patent Application No. PCT/US22/81153.