I'm working on a project where I have images annotated across several attributes, say
Z. Each image is also annotated with an attribute
M that is supposed to be an overall measure of a property in the image. This property
M is based on the different values of attributes
Z. I want to build a model that can predict a value for
M based on features that the model learns about
Z. My first thought was to build an ensemble model:
Model for X --| | Model for Y --|--- X_f * Y_f * Z_f --> Dropout -> Linear Layer -> ReLU -> Linear Layer -> Output | Model for Z --|
Essentially, I would train different models to predict values for
Z, and then I would stack these models into an Ensemble model. The same input would be passed in to each of the three models, and then I would take the feature maps the models produce before the last prediction layer, multiply them, and then pass them through a Dropout Layer, Linear Layer + ReLU, and a final Linear layer to predict
M. Is this a viable approach or is there a better way to handle this sort of ensemble modeling?
Most Ensemble model papers I've seen use different model architectures to predict the same thing, but my three models are predicting different variables. Any help would be appreciated!