Optimization of Neural Network Models Based on Symbol Interval Propagation

Node fusion is an important technique in neural network large model optimization. This technique can reduce the model size and accelerate the training and validation process. However, the fusion process is often accompanied by errors, resulting in a decrease in model reliability. This paper utilizes the symbol interval propagation method to derive the value intervals of nodes. And similar nodes are selected based on these intervals for fusion, with the objective of reducing errors and diminishing the size of the neural network model. Additionally, the intervals of nodes serve as a benchmark for quantifying errors in the fusion process. The experimental results show that this method can effectively reduce the fusion error and speed up the model optimization.