What should be the neural network response (output)?

Question

Neural networks

Hello, please tell me, I looked at various examples of neural networks, in them, as a rule, the output is YES (1) or NO (0), well, or 1 and -1, but let's say if I want to predict the temperature outside, or exchange rates, those. I have to get some specific value, let's say today is 20 degrees, and tomorrow it should be 24 and I need the neural network to return these 24 degrees to me.
Correctly, I understand that I am giving data like t, t-1, t-2, t-3 tn in a normalized form as parameters, and I should receive t + 1 as an answer.
Another stupid question from the previous question, should the answer also be in a normalized form? which I then need to translate into a normal form?
And the question is, if I get more than one answer, let's say for 3 days, i.e. 3 output parameters t + 1, t + 2, t + 3, I read somewhere that it is better if the network gives one answer, and when it works a lot worse, is it?
You can only please in simple language, on your fingers. Thanks a lot.

Answer 1

And how the people were fooled by neural networks.
The task that you described - predicting the temperature value, for example - is quite calmly solved by much more "human" methods, in particular - in the classical analysis of time series.
Tip - start learning ML - from the beginning (with probability, hypothesis analysis, regression, etc.), and not from the end (neural networks and Deep Learning).

Answer 2

An Aeroflot plane is flying.
The stewardess asks the passenger: - Will you eat?
He: What choice do I have?
Ta: Yes or no!

you have a mess in your head, start with, at least, this (I don’t remember if there is about automatic machines there or not, but at least something)
you need to understand, and not cramming, well, where does it fit:
all the network does is report the probability of a certain event:
you: will the temperature be 24 degrees tomorrow?
net: yes
you: what is the difference between duck
net: ...

Answer 3

1. The range of the output value depends on the activation function of the output. If we need a binary value of 0/1 - use sigm. If you need a non-negative (when we encode a color, for example) - use relu. If you just need a real one, then no activation function is needed. If you need a real value in some range (for example, -50/+50), you can try tanh and simply multiply it by 50 additionally.
2. Data normalization is useful, but not required. In the case of temperature, you can, for example, try to set not absolute values, but relative ones - how it has changed compared to the previous day.
3. To work with a data sequence (t-3, t-2, ...), a specific architecture of neural networks is usually used - recurrent neural networks, which for each new t recalculate some internal state and give a new value based on this state and the input value. The simplest thing you can do to make such a network give several predictions t+1,t+2,t+3 is to try to give it empty inputs t,t+1,t+2. Although it is not certain that it will work. Usually, in order to generate another sequence from a data sequence, the sequence-to-sequence architecture, seq2seq, is used.
4. The greater the variability of the output data; the more complex the function that the neural network needs to learn, the more difficult / longer / less accurate the learning itself will be, of course.