One substantial question, that is often argumentative in learning theory, is how to choose a `good' loss function that measures the fidelity of the reconstruction to the original. Logarithmic loss is a natural distortion measure in the settings in which the reconstructions are allowed to be `soft', rather than `hard' or deterministic. In other words, rather than just assigning a deterministic value to each sample of the source, the decoder also gives an assessment of the degree of confidence or reliability on each estimate, in the form of weights or probabilities. This measure has appreciable mathematical properties which establish some important connections with lossy universal compression. Logarithmic loss is widely used as a penalty criterion in various contexts, including clustering and classification, pattern recognition, learning and prediction, and image processing. Considering the high amount of research which is done recently in these fields, the logarithmic loss becomes a very important metric and will be the main focus as a distortion metric in this thesis. In this thesis, we investigate a distributed setup, so-called the Chief Executive Officer (CEO) problem under logarithmic loss distortion measure. Specifically, agents observe independently corrupted noisy versions of a remote source, and communicate independently with a decoder or CEO over rate-constrained noise-free links. The CEO also has its own noisy observation of the source and wants to reconstruct the remote source to within some prescribed distortion level where the incurred distortion is measured under the logarithmic loss penalty criterion. One of the main contributions of the thesis is the explicit characterization of the rate-distortion region of the vector Gaussian CEO problem, in which the source, observations and side information are jointly Gaussian. For the proof of this result, we first extend Courtade-Weissman's result on the rate-distortion region of the discrete memoryless (DM) $K$-encoder CEO problem to the case in which the CEO has access to a correlated side information stream which is such that the agents' observations are independent conditionally given the side information and remote source. Next, we obtain an outer bound on the region of the vector Gaussian CEO problem by evaluating the outer bound of the DM model by means of a technique that relies on the de Bruijn identity and the properties of Fisher information. The approach is similar to Ekrem-Ulukus outer bounding technique for the vector Gaussian CEO problem under quadratic distortion measure, for which it was there found generally non-tight; but it is shown here to yield a complete characterization of the region for the case of logarithmic loss measure. Also, we show that Gaussian test channels with time-sharing exhaust the Berger-Tung inner bound, which is optimal. Furthermore, application of our results allows us to find the complete solutions of three related problems: the quadratic vector Gaussian CEO problem with \textit{determinant} constraint, the vector Gaussian distributed hypothesis testing against conditional independence problem and the vector Gaussian distributed Information Bottleneck problem. With the known relevance of the logarithmic loss fidelity measure in the context of learning and prediction, developing algorithms to compute the regions provided in this thesis may find usefulness in a variety of applications where learning is performed distributively. Motivated from this fact, we develop two type algorithms: i) Blahut-Arimoto (BA) type iterative numerical algorithms for both discrete and Gaussian models in which the joint distribution of the sources are known; and ii) a variational inference type algorithm in which the encoding mappings are parameterized by neural networks and the variational bound approximated by Monte Carlo sampling and optimized with stochastic gradient descent for the case in which there is only a set of training data is available. Finally, as an application, we develop an unsupervised generative clustering framework that uses the variational Information Bottleneck (VIB) method and models the latent space as a mixture of Gaussians. This generalizes the VIB which models the latent space as an isotropic Gaussian which is generally not expressive enough for the purpose of unsupervised clustering. We illustrate the efficiency of our algorithms through some numerical examples.