Nonparametric enrichment in computational and biological representations of distributions

This thesis proposes nonparametric techniques to enhance unsupervised learning methods in computational or biological contexts. Representations of intractable distributions and their relevant statistics are enhanced by nonparametric components trained to handle challenging estimation problems. The first part introduces a generic algorithm for learning generative latent variable models. In contrast to traditional variational learning, no representation for the intractable posterior distributions are computed, making it agnostic to the model structure and the support of latent variables. Kernel ridge regression is used to consistently estimate the gradient for learning. In many unsupervised tasks, this approach outperforms advanced alternatives based on the expectation-maximisation algorithm and variational approximate inference. In the second part, I train a model of data known as the kernel exponential family density. The kernel, used to describe smooth functions, is augmented by a parametric component trained using an efficient meta-learning procedure; meta-learning prevents overfitting as would occur using conventional routines. After training, the contours of the kernel become adaptive to the local geometry of the underlying density. Compared to maximum-likelihood learning, our method better captures the shape of the density, which is the desired quantity in many downstream applications. The final part sees how nonparametric ideas contribute to understanding uncertainty computation in the brain. First, I show that neural networks can learn to represent uncertainty using the distributed distributional code (DDC), a representation similar to the nonparametric kernel mean embedding. I then derive several DDC-based message-passing algorithms, including computations of filtering and real-time smoothing. The latter is a common neural computation embodied in many postdictive phenomena of perception in multiple modalities. The main idea behind these algorithms is least-squares regression, where the training data are simulated from an internal model. The internal model can be concurrently updated to follow the statistics in sensory stimuli, enabling adaptive inference