Understanding the encoder-decoder paradigm

In Chapter 5, Time Series Forecasting as Regression, we saw that machine learning is all about learning a function that maps our inputs to the desired output:

Adapting this to time series forecasting (considering univariate time series forecasting to keep it simple), we can rewrite it as follows:

Here, t is the current timestep and N is the total amount of history available at time t.

Deep learning, like any other machine learning approach, is tasked with learning this function, which maps history to the future. In Chapter 11, Introduction to Deep Learning, we saw how deep learning learns good features using representation learning and then uses the learned features to carry out the task at hand. This understanding can be further refined to the time series perspective by using the encoder-decoder paradigm.

Like everything in research, it is not entirely clear when and who proposed...