While studying the field of deep learning, the questions I had from the beginning have still not been resolved.

In general, supervised learning is known to solve problems by comparing the results output from the data with the labels because the data is labeled.

In contrast, unsupervised learning is understood as analyzing patterns in input data without separate labels and outputting appropriate results for those patterns.

I have a question.

Autoencoders are commonly known as unsupervised learning models.

For example, when a specific vector is used as an input to an encoder model and there is a model in which the corresponding image is generated through a decoder, a loss function is often constructed by comparing the generated image with the correct image for learning.

So, I wonder if this should be classified as supervised learning at the stage of comparing the output results with the correct answer data.

In addition, when considering the transformer model for translation, the translated text output through the decoder is compared with the actual correct translation text to construct a loss function. I am curious as to whether this case also has a tendency for supervised learning.