Neural Substitution for Branch-Level Network Re-parameterization

Seungmin Oh, Jongbin Ryu

Abstract: We propose the neural substitution method for network re-parameterization at the branch-level connectivity. This method learns different network topologies to maximize the benefit of the ensemble effect, as re-parameterization allows for the integration of multiple layers during inference following their individual training. Additionally, we introduce a guiding method to incorporate non-linear activation functions into a linear transformation during re-parameterization. Because branch-level connectivity necessitates multiple non-linear activation functions, they must be infused into a single activation with our guided activation method during re-parameterization. Incorporating the non-linear activation function is significant because it overcomes the limitation of the current re-parameterization method, which only works at block-level connectivity. Restricting re-parameterization to block-level connectivity limits the use of network topology, making it challenging to learn a variety of feature representations. On the other hand, the proposed approach learns a considerably richer representation than existing methods due to the unlimited topology, with branch-level connectivity, providing a generalized framework to be applied with other methods. We provide comprehensive experimental evidence for the proposed re-parameterization approach. Our code is available at https://github.com/SoongE/neural_substitution.

External IDs:doi:10.1007/978-981-96-0966-6_7