Fig. 9. Architecture of the used Neural Network.

For both regression and classification, the network first consists of three stacked long short-term memory (LSTM) layers⁹⁸ of sizes 128, 128 and 64. For regression, the last LSTM is directly fully connected into the output layer returning a mean μ and variance σ, while for classification the output layer is preceded by another fully connected layer of size 20. The architecture is inspired by the successful applications of recurrent neural networks during the AnDi-Challenge^62,68,71,74.