Abstract: “Recently, a novel method, called the “least squares auto-tuning”, which can find hyper-parameters in LS problems (that minimize another (true) objective), is proposed [1]. Although nonconvex and cannot be efficiently solved, this problem can be approximately solved using a powerful proximal gradient method to find good hyper-parameters (for LS problems). In this short paper, to evaluate the effectiveness of the LS auto-tuning method, we realize numerical experiment on a classification problem using the Kuzushiji-MNIST dataset [2].”

Keywords: Least squares; hyper-parameter optimization; Kuzushiji-MNIST