Compressing LSTM Networks by Matrix Product Operators

摘要

We propose an alternative LSTM model to reduce the number of parameters significantly by representing the weight parameters based on matrix product operators (MPO), which are used to characterize the local correlation in quantum states in physics.