Fig. 3: Localization and detection performance.

a Contour plot of the intensity of the response of the 21 FBG transducers as a function of the position of the indentation showing that the sensitive regions of the transducers are significantly overlapped. b Temporal evolution of the force applied to the e-skin surface during indentation (top row) and spike trains of input layer neurons corresponding to the positive (middle row) and negative (bottom row) components of the FBG signals. c Instantaneous activity of the neurons in the output layer when an indentation is performed. d Temporal evolution of the applied force during an indentation (top row) and spike trains of the output layer neurons highlighted in subplot (c). The estimated position of the indentation is computed as the barycentre of the activities of the output neurons. e Medians of the localization error for the considered decoding strategies (blue: SF-DIR architecture; orange: SC-BIO architecture; light orange: three variants of the SC-BIO architecture: (i) without lateral inhibition (NO INH), (ii) with feedback inhibition replacing the original feedforward scheme (Feedback INH), and (iii) with both feedforward and feedback inhibition (Combined INH); gray: replication of the strategy described in ref. ²³ and based on standard paradigms of AI). Error bars represent standard deviations of the median error computed via blocked bootstrap (one-tailed Wilcoxon signed-rank test on spatially blocked localization errors: n = 120, t = 4804, p = 0.001). f Scatter plot of the localization error over the surface of the skin. g Median localization error for different degrees of weight quantization and different numbers of output neurons. Error bars represent standard deviations of the median error computed via blocked bootstrap. Best fitted parameters and measure of goodness of fit for the exponential model are reported in Supplementary Tab. S1. Subpanels (a), (b), (c), (d), (f) and (g) report the results for the SF-DIR architecture.