Supervised Learning Using Data-dependent Random Features with Application to Seizure Detection
The randomized-feature technique has been successfully applied to large-scale supervised learning. Despite being significantly more efficient compared to kernel methods in terms of computational cost, random features can be improved from generalization (prediction accuracy) viewpoint. Recently, it has been shown that such improvement can be achieved using data-dependent randomization. We recently proposed an algorithm based on a data-dependent score function that explores the set of possible random features and exploits the promising regions. The method has shown promising empirical success (on various datasets) in terms of generalization error compared to the state-of-the-art in random features. Restricting our attention to cosine feature maps, in this work, we provide exact theoretical constraints under which the score function converges to the spectrum of the best model in the learning class. We further present another application of the method in Epileptic Seizure Recognition.
