Generative model to generate hidden activations coming from a previously trained hidden layer

Question

I need to train a generative model to generate vectors which resemble the activations of a particular hidden layer of a neural network which has been previously trained. In particular, the hidden layer has been previously trained by merging outputs of a recurrent neural network and a convolutional neural network. Is it possible to use a standard Variational Autoencoder to do that? If not, is there a generative model which is better suited to generate them?

Thank you.

Answer 1

Yes it is possible. I don't think anyone will be able to tell you which generative model will work best, so you'll have to try it out for yourself. The paper Deep Predictive Coding Networks for Video Prediction and Unsupervised Learning might be of some interest to you, as they have a network which predicts activations, although not generatively.