We present our study of uncertainty quantification for SWIM networks through a Gaussian-process lens. We treat an ensemble’s hidden features as inducing an empirical feature kernel and ask how closely that kernel matches the target kernel. Two complementary targets are selected for this: kernel-family targets (RBF, Linear and Polynomial) and function targets, where the targets are rank-1 kernels from deterministic signals(e.g., cos 4πx, x2 ). On a shared subsample of inputs, we produce correlation-standardized Gram matrices. We measure the agreement of kernels via centered kernel alignment (CKA). We select a single feature-scalar parameter to run a lightweight pilot that maximizes CKA, then fix it for the final evaluation under full budgets. We compare three distributions: the baseline i.i.d. normal distribution, SWIM uniform variant, and supervised SWIM. We report the final correlation matrices via correlation heatmaps and their CKA scores.     «

We present our study of uncertainty quantification for SWIM networks through a Gaussian-process lens. We treat an ensemble’s hidden features as inducing an empirical feature kernel and ask how closely that kernel matches the target kernel. Two complementary targets are selected for this: kernel-family targets (RBF, Linear and Polynomial) and function targets, where the targets are rank-1 kernels from deterministic signals(e.g., cos 4πx, x2 ). On a shared subsample of inputs, we produce correlat...     »