DataPredict

General Guide For Neural Networks Settings

This guide assumes that you have basic understanding of neural networks. If not, you can find the resources online and have a look at how neural networks works.

Make note that this guide is an oversimplification and may not exactly cover the whole details of how neural network works.

Without further ado, let’s begin.

Layers

The choice of your layers are quite important. It is one of the factor determine the accuracy and the training speed of your neural network model.

Usually, it is recommended that you have few layers if you can determine what pattern leads to certain predictions. For example:

If you can determine the pattern, then I recommend you that you only build two layer neural networks as more complex models may produce less accurate outputs.

In other words, the more complex the pattern means more layers are needed to produce more accurate outputs (generally).

Here’s a formula for you to remember (for qualitative analysis, not quantitative):

Activation Functions

Different activation functions have different properties. It is very important to choose the correct ones to achieve high accuracy. Here are the functions with their properties listed below:

Weight Initialization

How we initialize our weights can affect how fast the model can learn. For example:

Regularization

Regularization avoids our model from “memorizing” the connection between the inputs and outputs, which could lead to lower accuracy. It ensure that the model generalizes the connection between the inputs and outputs.

Bias Neurons

The presence of bias neuron must not be underestimated. It allows the calculated values to move away from 0 instead of being centered to it. In most cases, these are usually added to each layer except the final layer.