Recurrent Networks for Short-Term Memory: Analysis and Solution - A Review

The author of this passage is discussing the potential of recurrent networks to store short-term memory using their feedback connections. They mention that this capability could be beneficial for various applications such as speech processing, non-Markovian control, and music composition. \n\nHowever, the author also acknowledges that the existing algorithms for learning what to store in short-term memory are not very efficient or effective, especially when there are long time lags between inputs and corresponding teacher signals. They state that these methods do not offer clear practical advantages over backpropagation in feedforward networks with limited time windows. \n\nThe passage concludes by mentioning that the article will provide an analysis of the problem and propose a potential solution to improve the learning and storage capabilities of recurrent networks.