AutoEncoder loss values

[MB-MuratBayraktar]

Hello everyone, I am getting a loss value of 0.95 and val_loss of 0.93 after training on P100 GPU for 50 epochs. The loss is not getting lower after that. After reading about the way AutoEncoders work, I understand that they learn the pattern of the normal data first, then they predict the abnormal behavior when they cannot reconstruct easily (that's when they are shown anomaly points). But when we have an unsupervised problem (no labels available), what can we do? AS I noticed, here in PyOD, we train on (X_train) without taking into consideration the separation of normal/abnormal when training. could this be the result of high loss values?

One small thing to note, when I plot the features according to the output labels after prediction, they cluster very well, which is so strange with such loss value.

Thank you

[yzhao062]

unsupervised autoencoder does not differentiate normal and abnormal samples during the learning process.
That being said, if training too much, then anomalies may be memorized by the network as well. Given #normal samples >> #anomalies, moderate training will make the network do better with normal samples.