In machine learning problems, loss function values for large datasets are often approximated stochastically. The most well-known of these techniques is stochastic gradient descent. But other stochastic approximations, such as stochastic quasi-newton methods, have been investigated (e.g. ). The matlab optimization toolbox does not include an explicit option for stochastic approximations to be made. However, the stochastic approximation can still be implemented in the body of the provided objective function. Has anyone done this? If so, can you report on the results?

Matlab Optimization Toolbox for Stochastic BFGS

alex

2017 3 13

0 个回答

13 次查看（30 天）

1 个投票

In machine learning problems, loss function values for large datasets are often approximated stochastically. The most well-known of these techniques is stochastic gradient descent. But other stochastic approximations, such as stochastic quasi-newton methods, have been investigated (e.g. https://arxiv.org/pdf/1401.7020.pdf).

The matlab optimization toolbox does not include an explicit option for stochastic approximations to be made. However, the stochastic approximation can still be implemented in the body of the provided objective function. Has anyone done this? If so, can you report on the results?