Recurrent Neural Network Theory:

Recurrent Neural Network Theory:

image from geeksforgeeks.

Are specifically designed for sequence data like

1. time series data i.e how many sales have happened over a previous time series or what the production has been over a previous time series.
2. Sentences
3. a sequence of words

Recurrent neural networks , you are applying it to sequence based data.

let's image a sequence:

[1, 2, 3, 4 ,5 ,6]

would you be able to predict a similar sequence shifted one time step into the future?

[2, 3, 4, 5, 6, 7]

so later we'll train a network to predict a time series, shift is over one time step into the future.And this would be really beneficial if we're trying to redict sales for the next day.

We can utilize recurrent neural network for this,

let's review how a normal neuron works in a feedforward network.

Normal neuron:

Just takes some input or multiple input and aggregate them by passing it through some sort of activation function. like sigmoid function and then have a output

Recurrent Neuron:

It's going to send the output back to it itself as input to it's neuron.

Input @ t-1

output @ t-1

|

V

Input @ t

output @ t

|

V

input @ t+1

output @ t+1

and so on.

This is also called unrolling recurrent neuron;

Important to note that these neurons is actually receiving both inputs from previous time step as well as inputs from current time step. So each of these neurons has 2 sets of input.

These cells that are function of input from previous time steps are also known as memory cells. Recurrent neural networks are also flexible in their inputs and output for both sequences and single vector values.