“Well” trained model of an Autoencoder NN does not match with predictions

Question

I’m using Keras API for training an 5-layers autoencoder with the following building:

• 84 input/ouput units
• 3 hidden layers, with [40,30,40] units respe