Fig. 1: Estimating latent coupling between image features and hippocampal spike entropy (PE).

a Recording sites of depth electrodes for all participants with available probe coordinates. Hippocampus sites in red, amygdala sites in blue (top: x = −21, middle: y = −19, bottom: z = −17). b: VGG16 (trained on Imagenet) was used to predict activation maps at five layers of varying depth (max pooling layers 1–5) for images previously shown to participants at encoding, resulting in feature-wise activation maps. The mean across layer-wise features is shown for two example images and max pooling layers 1, 3 and 5. We extracted three summary metrics per layer and feature (sum, standard deviation, number of non-zero elements) before subjecting each layer-wise summary matrix (# images * # features) to a principal component analysis (PCA). In all further analyses, we relied on the first component score of each image, layer, and summary metric. Example image used here from²⁰ (https://creativecommons.org/licenses/by/4.0/). c: Spike entropy was calculated per neuron and trial based on the first second of image encoding (for all neurons with PE > 0.0001 and trials with >1/3 of neurons spiking). In brief, permutation entropy works by transforming signals into patterns (here: length  = 3) and counting these patterns before calculating the Shannon entropy of the pattern distribution. d: Within-person correlations were computed by decomposing the rank-correlation matrix of trial-wise spike PE (per neuron) and image feature metrics using partial least squares (PLS). Singular value decomposition (SVD) of the rank-correlation matrix results in neural and stimulus weights per latent variable (LV). The weights of the first LV (first column outlined in black) were used to reduce the dimensionality of neural and feature matrices into scores for each trial. The rank correlation between both weighted variables represents the latent estimate of across-trial coupling between image features and hippocampus spike PE (right-most panel). Panels in (d) show data from a single subject in our sample.