Section: New Results

Machine Learning

In [19] , we proposed a a novel discriminative learning framework that estimates the parameters of a random walks segmentation framework using a training dataset. The main challenge we face is that the training samples are not fully supervised. Specifically, they provide a hard segmentation of the medical images, instead of a probabilistic segmentation. We overcome this challenge by treating the optimal probabilistic segmentation that is compatible with the given hard segmentation as a latent variable. This allows us to employ the latent support vector machine (LSVM) formulation for parameter estimation.

We developed several machine learning applications to fMRI data, including graph representations [25] and structured sparsity regularization [26] , [44] . A similar structured sparsity approach was applied in the development of a novel learning algorithm, the k-support regularized SVM, with applications to neuromuscular disease classification from diffusion tensor imaging [24] . Efficient training applications for taxonomic classification were developed in [21] , while a fine grained taxonomic image classification task was introduced in [45] . The role of non-maximal suppression in accurate and efficient object detection cascades was elucidated in [20] . A fast, consistent two-sample test based on kernelized statistics was developed in [33] .

In [34] , propose to address the misalignment of M/EEG samples by explicitly modeling time shifts of different brain responses in a classification setup. To this end, we use the LSVM formulation, where the latent shifts are inferred while learning the classifier parameters. The inferred shifts are further used to improve the signal-to-noise ratio of the M/EEG data, and to infer the chronometry and the sequence of activations across the brain regions that are involved in the experimental task.