Enhanced feature learning via regularisation: integrating neural networks and kernel methods (Q6887367)

This interesting paper studies enhanced feature learning via regularization integrating neural networks and kernel methods. To set the scene, in the case of high dimensional data analysis, feature selection methods are important. In this regard, in order to extract meaningful information from complex and high dimensional data sets, representation learning often aims to automate this process. In high dimensions, however, non-parametric methods often struggle which makes multi-index models, a good alternative method. The authors propose a new method for feature learning and function estimation in supervised learning via regularised empirical risk minimisation considering functions as certain expectations of some smooth function classes over all possible one dimensional projections of the data. The authors, in particular, construct a gradient-based computational method for the above suitably scaled expectation using kernel neural networks for particles to approximate the above expectation, where the positive homogeneity of the Brownian kernel leads to improved robustness to local minima. Numerical experiments support the authors complexity estimates given in the paper.\N\NThe paper is well written with a good set of references.