Parameter-Wise Sensitivity Fields: Semantic Parameter Similarity in Computational Graphs

This research note compares the semantic similarity of parameters in a neural network (or general computational graph) by constructing parameter-specific sensitivity matrices. Each matrix records the partial derivatives of all internal and output states with respect to a single parameter. The matrices align with the graph’s node structure. Collecting these matrices produces sensitivity fields over the graph, capturing the functional influence of each parameter. Comparing parameters via similarity metrics over the fields enables functional clustering, interpretability, and detecting redundancies without retraining.