Wasserstein Distance Transfer Learning Algorithm based on Matrix-norm Regularization

Wasserstein distance has been applied in the transfer learning algorithm, but the existing methods are not ideal for the solution of Lipschitz constraint condition. To solve this problem, a new transfer learning method, namely Wasserstein distance transfer learning algorithm based on matrix-norm regularization (MRWDTL), is proposed. To satisfy the Lipschitz constraint, we process the weights of some network layers to propose the matrix-norm regularization term. We then implement this term as a regularization term in the loss function in order to learn domain invariant features using the Wasserstein distance. Through experiments on public datasets, it is demonstrated that MRWDTL can achieve superior results compared to some existing methods.