Probabilistic inference serves as a popular model for neural processing. It is still unclear, however, how approximate probabilistic inference can be accurate and scalable to very high-dimensional continuous latent spaces. Especially as typical posteriors for sensory data can be expected to exhibit complex latent dependencies including multiple modes. Here, we study an approach that can efficiently be scaled while maintaining a richly structured posterior approximation under these conditions. As example model we use spike-and-slab sparse coding for V1 processing, and combine latent subspace selection with Gibbs sampling (select-and-sample).

Overall this is an interesting and fairly clear algorithmic and experimental paper paper presenting an approximate EM technique for dictionary learning in a Bernoulli-Gaussian setting. The main idea is that EM requires sampling to approximate certain intractable expectations/integrals, and that such sampling is inaccurate or difficult in high dimensions H. It is proposed to perform sampling only after a selection step which consists in finding a subspace (dependent on the considered training vector as well as the currently estimated parameters of the model) of intermediate dimension H' H. As far as I could check the selection step, expressed implicitly in (10)-(11) corresponds to computing certain weighted correlations of the training sample y_n with columns of the current estimate of the dictionary, and to keep the H' largest correlations. This essentially boils down to one step of hard thresholding, reminiscent of certain recent techniques for dictionary learning (see, e.g., recent work of K. Schnass such as [A]). Numerical experiments on synthetic data show state of the art performance of the proposed approach.

Probabilistic inference serves as a popular model for neural processing. It is still unclear, however, how approximate probabilistic inference can be accurate and scalable to very high-dimensional continuous latent spaces. Especially as typical posteriors for sensory data can be expected to exhibit complex latent dependencies including multiple modes. Here, we study an approach that can efficiently be scaled while maintaining a richly structured posterior approximation under these conditions. As example model we use spike-and-slab sparse coding for V1 processing, and combine latent subspace selection with Gibbs sampling (select-and-sample).