Why MATLAB uses the Newton’s Gradient Descent as the solver for fitcnet?

I want to train a neural network to perform image segmentation using fitcnet. I noticed that MATLAB uses the Newton’s Gradient Descent solver (LBFGS) to perform parameters optimization. Does anyone know why MATLAB selected this method over other optimization methods such as Adam or SGDM?

I would appreciate any help you could provide.

Answer 1

[Lucas García]

L-BFGS is used both in in fitcnet and fitrnet. These functions ship in Statistics and Machine Learning Toolbox and allow you to get started in solving Machine Learning problems using Neural Networks.

For more customization capabilities, more advanced architectures and additional solver methods, you may use Deep Learning Toolbox. This includes the use of solvers such as SGDM, RMSProp or Adam. See trainingOptions for more details.

[Walter Roberson]

This does not answer the "why?" question however. Why was one chosen over the other.

... the answer might, for example, be "performance". You do not always need the extra complexity, so it can make sense to make the extra complexity something you can enable.

[Lucas García]

Good point Walter. I wanted to highlight that the solver available in Statistics and Machine Learning Toolbox is more appropriate for shallow networks (i.e. MLPs). L-BFGS has great convergence properties for small networks using smaller datasets.

For more complex networks, such as the one highlighted in the question (image segmentation), Deep Learning Toolbox does provide solvers such as SGDM and Adam (more appropriate for deeper network architectures).

It might be worth checking one of the Computer Vision Toolbox examples leveraging Deep Learning Toolbox: https://www.mathworks.com/help/vision/ug/semantic-segmentation-using-deep-learning.html

[Memo Remo]

Dear Lucas and Walter,

Thank you very much for your responses. They are very helpful.