I have already a feed-forward neural network model in Tensorflow.Keras called model and I found some methods to take the derivative of the output based on the model inputs. These are example posts for taking the derivative of a neural network method based on its input, post1 and post2.
Now, my question is going further by knowing how to find an input vector that minimizes the output of a neural network with only one output?
Although a possible solution is using some black-box optimization techniques to find the minimum point, I am wondering can it be possible to solve it symbolically somehow, e.g., gradient = 0, based on the obtained derivative explained in reference posts?
