Can we use different architectures of Generative Adversarial Network (GAN) to solve P vs. NP problem by Artificial Intelligence. For example, in case of Zeroes of Riemann Zeta Function, How can we modify the semantic of discriminator of GAN for not converging to 1/2 in real part?
No. And the question, whether there exists an algorithm that can solve any problem in a time that's a polynomial in the size of the problem doesn't have anything to do with the location of the zeros of the zera function.
It ought to be well known, incidentally, that, while any given neural network can describe a number of associations that's exponential in the number of the units, the neurons, it can represent only a number of associations that's a linear function of the number of neurons.
Stam Nicolis : Calculating the location of zeroes of Zeta Function was an example of problem which I referred. As using general GANs to find zeroes with non 1/2 real part is non-achievable since the discriminator aligns itself to 1/2 and rejects the possibility of solutions other than 1/2 real part.
Hence, Can the non-linear activation function of neurons in ANNs/CNNs in Generative Adversarial Network (GAN) be modified to solve P vs. NP problem such as the problem described above?
The short answer is No: The fact that a GAN can represent an association doesn't have anything to do with whether representing it is a problem whose complexity necessarily grows in a polynomial way or doesn't.
On the other hand, whether all the zeros of the Riemann zeta function have real value 1/2 isn't a statement that can be expressed any GAN, that's why it doesn't make sense looking for one at all. There's simply no association to be learned.
- Badges
- Science topic
- Similar topics
- Condensed Matter Physics
- Zero