Background: Recent functional magnetic resonance imaging (fMRI) decoding techniques allow us to predict the contents of sensory and motor events or participants' mental states from multi-voxel patterns of fMRI signals. Sparse logistic regression (SLR) is a useful pattern classification algorithm that has the advantage of being able to automatically select voxels to avoid over-fitting. However, SLR suffers from over-pruning, in which many voxels that are potentially useful for prediction are discarded. New method: We propose an ensemble solution for over-pruning, called "Iterative Recycling" (iRec), in which sparse classifiers are trained iteratively by recycling over-pruned voxels. Results: Our simulation demonstrates that iRec can effectively rectify over-pruning in SLR and improve its classification accuracy. We also conduct an fMRI experiment in which eight healthy volunteers perform a finger-tapping task with their index or middle fingers. The results indicate that SLR with iRec (iSLR) can predict the finger used more accurately than SLR. Further, iSLR is able to identify a voxel cluster representing the finger movements in the biologically plausible contralateral primary sensory-motor cortices in each participant. We also successfully dissociated the regularly arranged representation for each finger in the cluster. Conclusion and comparison with other methods: To the best of our knowledge, ours is the first study to propose a solution for over-pruning with ensemble-learning that is applicable to any sparse algorithm. In addition, from the viewpoint of machine learning, we provide the novel idea of using the sparse classification algorithm to generate accurate divergent base classifiers.
Hirose, S., Nambu, I., & Naito, E. (2015). An empirical solution for over-pruning with a novel ensemble-learning method for fMRI decoding. Journal of Neuroscience Methods, 239, 238–245. https://doi.org/10.1016/j.jneumeth.2014.10.023