The idea of "self-network" is to use the potentials inside the optical components as weights, and apply backward propagation in the similar way as classical neural networks -> no need external neural network to calibrate the optical component, it will act itself as the neural network, with the weights being what we are searching for.
The idea of "self-network" is to use the potentials inside the optical components as weights, and apply backward propagation in the similar way as classical neural networks -> no need external neural network to calibrate the optical component, it will act itself as the neural network, with the weights being what we are searching for.