Reservoir Transfer Learning Through Source Enhancement Reconstructed With Reservoir Stein Discrepancy
Jia-Wei Lin, Fu-Lai Chung, Shitong Wang · IEEE Internet of Things Journal · 2024
Due to the use of reservoir states, reservoir transfer learning may perhaps face an overwhelming phenomenon—-possibly more dissimilarity between reservoir states, respectively, from the original source and target domains. This phenomenon inevitably triggers three challenges: fast calculation of discrepancy between the reservoir source and target domains; reservoir source enhancement through reconstruction; efficient transfer learning on the reconstructed source domain. In this study, based on the well-known leaky integrator echo state network (LI-ESN), an efficient reservoir transfer learning method is developed to battle the above challenges. With the theoretical justification, the proposed method has three steps. The first step realizes fast calculation of the manipulatable reservoir Stein discrepancy metric for quantitatively measuring the discrepancy between the known reservoir source distribution and the unknown reservoir target distribution. The second step is in charge of reservoir source enhancement through state reconstruction for the generation of more similar reservoir source domain from the target domain. And the third step accomplishes an efficient target classifier by realizing LI-ESN based reservoir transfer learning method on both the reservoir target domain and the reconstructed reservoir source domain. Experimental results on a heap of the adopted datasets indicate that the superiority of our proposed method when facing the above three challenges under reservoir transfer learning scenarios.