Test different approaches to improve the score

[GiovanniPioDelvecchio]

It is necessary to test different approaches to improve the current best model. To do so it is possibly to experiment:

• adding convolutions (1D) before the LSTM layers.
• add a batch normalization layer after the above mentioned convolutions.
• add max-pooling after the batch normalization.
• try ELU instread of ReLU. Source: https://www.mdpi.com/1424-8220/22/11/4157

[DomMcOyle]

other enhancement to be tested are:

• add padding masking to the best performing model so far (the one classifying only the cell states from the lstm)
• retry the "semeraro" head without adding the concatenation and reducing the bottleneck introduced by the max pooling operations