An alternative approach to train neural networks using monotone variational inequality