Autoencoder feature extraction without validation set?

Question

I plan to use autoencoder for feature extraction, then use the latent vector for clustering.

My autoencoder performs very very well on my training set (loss small and reconstructed image look very similar to input) but shows lower performance on validation set.

Is it a good idea to NOT use validation set but only a training set? Since I can reach good performance extracted vector should be "good" features

If I use new data I'll need to retrain the model, but I should always be able to reach good performances since new data will be very similar to current data

Answer 1

To the best of my knowledge, if the performance is higher on the training set than on the validation set, you should probably be worrying about over-fitting.

I am assuming that your validation set, for the moment, is just a fraction of the data of the training set.

Notice that if the autoencoder is not performing well on your validation set, it is unlikely that you will reach good performance on new data "since new data will be very similar to current data", as already on current data you don't have good performance.

If you have to retrain the autoencoder every time you want to encode something, it may be not worth using the autoencoder in the first place.