Question of End-to-end Learning with Autoencoder

[Mio-TW]

I have the question in https://nvlabs.github.io/sionna/examples/Autoencoder.html#Trainable-End-to-end-System:-RL-based-Training

For the RL-based Training, why "-" is added in the tx_loss? It may cause the loss become negative value.
If there is no negative sign, increasing the power of "p" will result in a larger BCE and consequently a larger tx_loss. However, a negative sign is added, the outcome become opposite. I am curious about the reason behind this.

Thank you ao much.

[faycalaa]

Hi,

The code you are referring to implements the algorithm from this paper and more specifically equation (8) in the case where \pi corresponds to the density of a normal distribution. Taking the log of this density results in the implemented equation, where the "-" corresponds to the one that is in the exp(.) of the density.