I want to train a shallow neural network using known output gradients rather than input/output training pair data

Question

James 2022-7-19

回答： arushi 2024-8-27

I have a shallow network defined by:

net = fitnet([64,112],'traingd');

The outputs of this network feed INTO a function Y = f(X) where X is the vector of net outputs X=net(I).

I calculate the gradient of Y w.r.t X and want to then train net based on these gradients rather than input/output data for the net.

Answer 1

arushi 2024-8-27

Hi James,

I understand that you are trying to train a shallow neural network using the known output gradients.

To train a shallow neural network using known output gradients, you can follow these steps:

Initialize your shallow neural network with the desired architecture.
Define your function Y = f(X) that takes the network outputs X as input and produces the desired output Y. This function represents the relationship between the network outputs and the target outputs.
Calculate the gradient of Y with respect to X.
Use the calculated gradients to update the network weights using gradient descent or another suitable optimization algorithm.

For additional information, please refer to the following documentation:

I hope this helps!