RNNs work extremely well at solving problems involving processing text, video, and audio data, mostly because of the arbitrary data input length. But why are they that good at working with time series?
The point is, that RNNs can hold a state, that basically means that the model can learn which information it wants to save and what to get rid of judging on ordering and the design of creation and passing of the state.
It means, that RNN is able just to use the previous state and process on a new singular input or a smaller context window?
Yes. In RNNs, the output of the previous state is passed as an input to the current state. On a hunch, there is a time-based relationship in the way input is processed in RNN. The net can understand how the current state was achieved on the basis of the previous values. Like, a value at time-step x is a result of value at time-steps x-1, x-2 and such.
Just drop us an email to ...
Show more 
Log in with Google