3 Mixture of Factor Analyzers based Generative Model for Multitask Learning

We propose a general framework for learning multiple related learning tasks. The proposed unified framework can capture various types of latent structures underlying the weight vectors of multiple tasks. In doing so, our model can automatically interpolate to the appropriate task relatedness assumption as warranted by a given dataset. For instance, the model can capture multitask learning notions such as a shared Gaussian prior in the parameter space, task clustering, lowrank assumption, etc. as special cases, or adapt itself to a more general combination of these assumptions, addressing their individual shortcomings. Our model, therefore, brings in considerable flexibility as compared to these commonly used multitask learning models that are based on some a priori fixed notion of task relatedness. We also present an efficient inference algorithm for this model. Experimental results on several real-world datasets, on both regression and classification problems, establish the efficacy of the proposed method. νθt123 2 1μνθt