This Repository implements a pytorch version of the LRU, including parallel training and serial inference.
The parallel acceleration is achieved by Upper/Lower algorithm based on Bojbojone/rnn.
LRU is proposed by Resurrecting Recurrent Neural Networks for Long Sequences as an alternative to RNNs, with an architecture inspired by SSMs. I conducted experiments on a classification problem in finance, and the results showed a significant improvement of LRU over LSTM.
If you want to use LRU for time series prediction, please note that the paper requires LRU with MLP for this, while merely LRU is implemented in this project.