How does Long Short-Term Memory (LSTM) architecture help neural networks retain useful information across long sequences? What makes LSTM different from a traditional RNN? How can LSTM models support tasks such as time series forecasting, language processing, and speech recognition?