Hessian-based Bounds on Learning Rate for Gradient Descent Algorithms
Prayag Gowgi, Shayan Srinivasa Garani · 2020
Learning rate is a crucial parameter governing the convergence rate of any learning algorithm. Most of the learning algorithms based on stochastic gradient descent (SGD) method depend on heuristic choice of learning rate. In this paper, we derive bounds on the learning rate of SGD based adaptive learning algorithms by analyzing the largest eigenvalue of the Hessian matrix from first principles. The proposed approach is analytical. To illustrate the efficacy of the analytical approach, we considered several high-dimensional data sets and compared the rate of convergence of error for the neural gas algorithm and showed that the proposed bounds on learning rate result in a faster rate of convergence than AdaDec, Adam, and AdaDelta approaches which require hyper-parameter tuning.