Mentions a so called HSIC bottleneck, maximizing mutual information between hidden states and labels, minimizing mutual dependence between input and hidden stat

Mentions a so called HSIC bottleneck, maximizing mutual information between hidden states and labels, minimizing mutual dependence between input and hidden states (?) PDF here https://arxiv.org/pdf/1908.01580v1.pdf https://arxiv.org/pdf/1908.01580v1.pdf Another place I recently saw Reproducing Kernel Hilbert Spaces was in Stein Variational Gradient Descent (PyMC3 implements it). It seems like a clever basis set but have to admit I don't understand what's going on.