Is the output value in the range neural network greater than 1 or less than -1?

Question

Neural networks

A week ago, it became interesting to study the work of neurons. To date, I was able to write some simple network like a perceptron. Used the back propagation method by finding the partial derivative + sigmoid to activate neurons.

Then it became interesting for me to write the recognition of numbers in the picture (i.e., according to the input parameters of the pixels, but as it turned out later, it’s worth studying a separate situation for recognizing individual images (correct if I misunderstood this)).

That's why I decided to write a forecast of time schedules. I thought it would be enough to simply predict the graph of some kind of sinusoid for a neuron, because in fact, you only need to input the y-axis (one neuron) and several hundred training data. However, there was a problem with the fact that all the data inside the HH is normalized by the activation function, so I can only get a number from 0 to 1 at the output (1 to -1 if I use tahn).

How can you get the output value of a neuron in a wider range? For example, I want to know the approximate price of a product, according to some input data (type of product, quality of packaging, etc.). I've tried disabling normalization entirely, but for some reason it crashes due to too large numbers. All my neurons are randomly generated from 0 to 1 before training.

Answer 1

You can use a linear activation function on the output neuron, then the numbers can be any. But it usually doesn't end well.