Probabilistic Implicit Multitask Feature Learning Based on Weight Matrix Partition
Yu Zhang, Xin Zuo, Jian–wei Liu · Journal of Physics Conference Series · 2018
Multitask learning (MTL) can be considered at two levels, considering the multiple features attribute that can be migrated between the most tasks simultaneously called MTL on task level, and only considering the features attribute that can be migrated between a very few tasks called MTL on feature level. A single feature is composed with an overlying structure including these above two levels, since the similarity between the multi-tasks is not unique. In the previous MTL studies, in task level, the shared feature structure of inter-task is the leading approach, and in feature level, only considering a partial of salient features is more prevalent. However, this model and structure assumptions are violated on most complex learning scenarios, and some specific feature structures is also limited. In this paper, from the Bayesian viewpoint, we propose a new method, the weight matrix is defined by stochastic variables which probabilistic distribution is overlaid, the relationship in task level are no longer represented by a unified structure, and relationship in feature level are not limited to concrete structural forms. The overlying structure is defined by two parts: multidimensional Gaussian mixture distribution according to the overall features to describe the task relationship and one-dimensional Gaussian mixture distribution to describe the feature relationship on every feature. These two parts can be elaborated by the information contained in each Gaussian sub-classes separately, then we determine the closeness of the relationship by calculating the posterior probability distribution. The advantage of our method is that we can operate the weight matrix more flexible and make full use of the probability distribution to get abundant expression in the task level and grasp the key factors in the feature learning process.