Structured sparsity model with spatial similarity regularisation for semantic feature selection

In the prediction of brain activity associated with concrete concepts, the main task is to construct a computational model to reveal the neural basis of the concepts. However, the ordinary regression model cannot select desired semantic features and easily over-fitting. To address these problems, in this paper, we propose a structured sparsity model to automatically choose the relevant semantic features by exploiting the sparsity of responses and the spatial relationships between the voxels. Specifically, we require the number of the non-zero responses to be sparse and the responses that two voxels are nearby in the brain to be similar. The constraints do not only regularise the model fitting but also have an interpretation in terms of brain hemodynamics. The experimental results on predicting brain images show the effectiveness of the proposed approach, as well as improved interpretability.