A Novel Utility-Driven Residual Connection Inspired by Debreu's Theorem

This paper proposes a residual connection mechanism based on the idea of Debreu's theorem in economics and applies it to the multi-layer perceptron model (MLP). By introducing a "utility function" module with monotonicity constraints as an auxiliary channel, we add a structure that simulates preference enhancement on the basis of the original feature map to form a "utility residual connection". Experimental results show that on the CIFAR-10 dataset, the improved model has a slight improvement in accuracy, recall rate and F1 score compared with the standard MLP, while the training time is slightly reduced. More importantly, the structure has good versatility and can be extended to convolutional neural networks, graph neural networks and transformer structures, providing a new inspiration path for the design of deep learning structures.