Adaptive neural activation functions in multiresolution learning

In this paper, we extend our original work on multiresolution learning, and present a new concept and method of adaptive neural activation functions in multiresolution learning, to maximize the learning efficacy of multiresolution learning paradigm for neural networks. Real-world sunspot series (yearly sunspot data from 1700 to 1999) prediction has been used to evaluate our method. We demonstrate that multiresolution learning with adaptive activation can further significantly improve the constructed neural network's generalization ability and robustness. Therefore, our work demonstrates the synergy effect on network learning efficacy through multiresolution learning with neural adaptive activation functions.