LOSSGRAD: automatic learning rate in gradient descent

Abstract

In this paper, we propose a simple, fast and easy to implement algorithm LOSSGRAD (locally optimal step-size in gradient descent), which automatically modifies the step-size in gradient descent during neural networks training. Given a function $f$, a point $x$, and the gradient ${\nabla }_{x}f$ of $f$, we aim to find the step-size $h$ which is (locally) optimal, i.e. satisfies: $$h=argmi{n}_{t\ge 0}f(x-t{\nabla }_{x}f).$$ Making use of quadratic approximation, we show that the algorithm satisfies the above assumption. We experimentally show that our method is insensitive to the choice of initial learning rate while achieving results comparable to other methods.

arXiv commentTFML 2019

References in corpus (8)

• Adam: A Method for Stochastic Optimization
• Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms
• ADADELTA: An Adaptive Learning Rate Method
• An overview of gradient descent optimization algorithms
• WNGrad: Learn the Learning Rate in Gradient Descent
• A Walk with SGD
• L4: Practical loss-based stepsize adaptation for deep learning
• Probabilistic Line Searches for Stochastic Optimization