Initializing LSTM internal states via manifold learning

We present an approach, based on learning intrinsic data manifold, for the initialization of internal state values long short-term memory (LSTM) recurrent neural networks, ensuring consistency with initial observed input data. Exploiting generalized synchronization concept, we argue that converged, “mature” states constitute a function this learned manifold. The dimension manifold then dictates length time series required consistent initialization. illustrate our approach through partially chemical model system, where initializing LSTM in fashion yields visibly improved performance. Finally, show enables transformation dynamics into fully ones, facilitating alternative identification paths nonlinear dynamical systems.