Sequence Data
Study of machine learning algorithms designed for sequential data is known as sequence learning
In a neural network, you have one output per input. You have an image and you have a label, for instance. There is no way you can input image after image on Neural Networks and get an output based on all of them. The nature of neural networks make them unable to process sequential data.
So, in a feed forward neural network, output at a time t is independent of the output at time (t-1).
On the other hand, RNNs work very well with sequential data. They have a mechanism of “remembering” the previous inputs and producing an output based on all of the inputs. This makes them well-suited for sequential type of data such as text, audio, video or any time-series data.
The brain consists of billion of neurons, without any single duration.
A Decision made now is not only based only on what you hear/see now.
We can think and reason based on past inputs.
What happens if we add feedback loops and memory to neural network.
Traditional neural networks cannot do this, and it seems like a major shortcoming. For example, imagine you want to predict the next word in a sentence given the previous words and the nature of the sentence. It would be impossible for a traditional neural network to predict the next word because it would not be able classify the options to get a word that would make the sentence meaningful. That's the property of the sequence data, where one should have a classifier that stores and uses that past data to create a model that predicts the future. Hence RNN models are the prefect example to use for forecasting sequence data.
Last updated
