Sensitivity Analysis of Neural Network through Meta-Learning

Abstract

The machine learning can provide high prediction accuracy, but the model sometimes becomes a black box, and there is a problem in the interpretability of the relationship between input and output of the model. PDP, ICE Plot, Forest Floor have been proposed as visualization for the model interpretation. The objective of this study is to investigate a model-independent sensitivity analysis. Meta-learning is a process of learning to learn intends to design model, and methods such as MAML and OpenAI Reptile have been proposed. We conducted an experiment using OpenAI Reptile for sensitive analysis of the neural network, and discussed the results.