Abstract:
|
As a consequence of the technological advancements in many modern applications, complex structured datasets are routinely collected with increasing efficacy. Their complexities are pushing the new developments of novel statistical modeling and inferential tools. One such scenario arises when data is generated from heterogeneous subpopulations. Hierarchical models form state-of-the-art methods for such scenarios. Instead of a multi-stage estimation procedure, a Bayesian route for inference is usually preferred. However, posterior approximation using Markov Chain Monte Carlo methods can become computationally prohibitive for such large-scale models due to the time complexities. Sometimes, they may also suffer from inconsistency issues. On the other hand, mean-field VI methods or MAP estimates, even though fast, can suffer from inefficiency due to non-convexity of resulting posterior. Examples of such situations for trade-offs between statistical and computational efficiencies occur in numerous other forms. The overarching goal of this roundtable discussion is to provide an extensive and broad understanding of such scenarios across modern Statistics and Machine Learning literature.
|