I am writing a paper on a model I build for analyzing a dataset. The model is an autoencoder with an additional unit attached to the embedding layer. Similar to an adversarial autoencoder, except, in this case, the new unit can train for any prediction; binary or multi-class classification, another decoder, regression model, LSTM, and so on. The embeddings feed into the new unit or units and train in a supervised style.
The training goes through multiple passes; based on the available input and output, it can feedforward in various paths. For example, if I use it with a classifier attached, the classifier weights will update if the labels are available. Otherwise, it will only update the encoder and decoder.
It’s doing a good job on my dataset. I have looked in the literature to find similar architectures and reference them in mine, but the most similar case I found is the adversarial autoencoder. I would appreciate it if anyone gave me feedback on this and let me know if they have read any papers with similar architecture.
Thanks.