Is it possible to perform inverse prediction using a neural network using the Matlab Deep Learning toolbox?

This function allows for forward prediction of a neural network: https://www.mathworks.com/help/deeplearning/ref/predict.html

Is it possible to perform inverse prediction where the response is input to the trained network and an image is output?

I know it is mathamatically possible, however, is there an easy way to do this using the Matlab Deep Learning toolbox? Thanks!

Answer 1

[Walter Roberson]

Swap responses and inputs and train on that. There is no inherent reason why the number of outputs needs to be larger than the number of inputs.

However, this would be used for generating an image given the responses, rather that matching to find out which of a set of known input images comes closest to generating the known responses.

[jman4162]

It is possible to swap the inputs and outputs and train again, however, this is not what I am trying to do. Once the neural network is training, I would like to be able to solve inverse design problems using the gradiatents calculated during backpropogation. Does the Matlab Deep Learning provide the functonality to use the numerical gradients for backwards calculations? Thanks!

[Walter Roberson]

Deep Learning Toolbox used to be named Neural Network Toolbox, when it supported just the standard network types such as feed forward or perceptron. It was renamed once it started getting heavily involved in Recurrent networks and gpu training and the like.

The facilities available for the newer network types work fairly differently than the older types, and I do not know much about how they are put together.

predict() as a function is from the older network types. As best I recall, the older network types do not support configuring gradients for back propagation. I would not want to say it is not possible for sure, as perhaps there are hacks such as using custom transfer functions.

Greg Heath is the volunteer specialist for the older NN but he has cut back on his contributions and might not happen to notice this question.

[jman4162]

Thank you very much for providing your insight.