Estimation of LAIS was performed using the open source Java information dynamics toolkit (JIDT) (Lizier, 2012 ), with a history parameter kmax of ten time points, spaced 2 samples, or (2/150 Hz)= 13.3 ms, apart. The total history length thus covered 133 ms, or roughly one cycle of a neural theta oscillation, which seems to be a reasonable time horizon for a downstream neural population that ultimately must assess these states. To enable LAIS estimation from a sufficient amount of samples, we considered the data pixels as homogeneous variables executing comparable state transitions, such that the pixels form a physical ensemble in terms of information storage dynamics. Pooling data over pixels thus enables an ensemble estimate of the PDFs in question. This approach seems justified as all pixels reported activity from a single brain area (area 18 of cat visual cortex, see below). Mutual information was estimated using a box kernel-estimator (Kantz and Schreiber, 2003 (link)) with a kernel width of 0.5 standard deviations of the data.
Here we assume that the neural system is at least capable of exploiting the statistics arising from the stimulation given throughout the experiment and thus construct PDFs from all data (time points and pixels) for a given condition. Therefore, we pool data over the full time course from −1 to 1 s of the experiment. Thus, each image of the VSD data had a spatial configuration of 67 × 137 spatial data pixels after removal of the two rows/columns on each side of an image because of the median filter that was applied. Each trial (of a total of eight trials per condition) resulted in 288 LAIS values, based on an original data length of 298 samples and a history length (state dimension) of 10 pixels. The product of final image size and LAIS samples resulted in 2.64 · 106 data points per trial for the estimation of the PDF for each of the eight motion direction conditions. Due to computational limitations, LAIS estimates were performed on two blocks of four trials separately, resulting in 1.06 · 107 data points entering the estimation in JIDT.
Here we assume that the neural system is at least capable of exploiting the statistics arising from the stimulation given throughout the experiment and thus construct PDFs from all data (time points and pixels) for a given condition. Therefore, we pool data over the full time course from −1 to 1 s of the experiment. Thus, each image of the VSD data had a spatial configuration of 67 × 137 spatial data pixels after removal of the two rows/columns on each side of an image because of the median filter that was applied. Each trial (of a total of eight trials per condition) resulted in 288 LAIS values, based on an original data length of 298 samples and a history length (state dimension) of 10 pixels. The product of final image size and LAIS samples resulted in 2.64 · 106 data points per trial for the estimation of the PDF for each of the eight motion direction conditions. Due to computational limitations, LAIS estimates were performed on two blocks of four trials separately, resulting in 1.06 · 107 data points entering the estimation in JIDT.
Free full text: Click here
