Poor results in the Deep Network Designer app

Hey!

I am doing a regression neural network with only numerical values (electricity price analysis) and any way I try to use the "Deep Network Designer" app gives me worse results than using the "fitnet" function when I use the same layers in the app as the ones used in the function. It is true that the training functions are different in both cases and a couple of other things but I don't understand why the difference is so abysmal.

I think this app is not very well thought for purely numerical values and is more focused on image classification.

Still, as the app has more design possibilities I would like to know how to increase the efficiency of this one.

Answer 1

[Himanshu]

Hello Fernando,

As per my understanding, you want to improve the efficiency of your regression model using the Deep Network Designer app.

The Deep Network Designer app in MATLAB provides an interactive environment for designing, analyzing, and training deep learning networks. While it offers more design possibilities and flexibility, setting up your network and training options correctly is essential to get the best results.

You can follow the steps below to increase your neural network's efficiency.

1. Preprocess your numerical data by normalizing the input features to ensure they have similar ranges, which can help the network learn faster and more efficiently.
2. Experiment with the network architecture to find the best one for your regression problem. Try adding or removing layers, changing the number of neurons in each layer, or using different activation functions. Since you're working with a regression problem, ensure the last layer has a single neuron with a linear activation function.
3. Adjust the training options in the Deep Network Designer app to match the options used in the "fitnet" function. Some key training options are solvers, initial learning rate, mini-batch size, maximum epochs, etc.
4. Experiment with different regularization techniques like L1 or L2 regularization, dropout layers, or early stopping to prevent overfitting.
5. Conduct a systematic search for the best hyperparameters using techniques like grid search, random search, or Bayesian optimization.

You can refer to the below documentation to understand more about Deep Network Designer app.

Design, visualize, and train deep learning networks - MATLAB - MathWorks India

Normalize data - MATLAB normalize - MathWorks India

Options for training deep learning neural network - MATLAB trainingOptions - MathWorks India

Regularization - MATLAB & Simulink (mathworks.com)

Tune Experiment Hyperparameters by Using Bayesian Optimization - MATLAB & Simulink - MathWorks India

[FERNANDO CALVO RODRIGUEZ]

Thanks for the answer, do you also know how to add cross validation or regularization to the "fitnet" function or how to add more than one final output in the "fitnet" function (this is another question I just asked).