Characterizing Neural Network Weights for Class Imbalanced Learning
R. Siddharth, Abhik Banerjee, Vijendran G. Venkoparao · 2023
Exploring the weight space of the neural network can lead to insights about how the network makes decisions. However, interpreting information within the weight space poses significant challenges, representing an active and complex area of ongoing research. This study is driven by the motivation to explore the hidden patterns in the weight space. The motivation of the study is to understand the weight space by characterizing the pattern space of the original and augmented samples, utilized for oversampling the underrepresented class in imbalanced learning. The major contribution of this paper is to address the fundamental task of highlighting the metrics that effectively characterize the pattern space. This characterization is the interim approach toward the motivation of our study regarding understanding the weight space of a neural network. Moreover, this research provides a comprehensive exploration of the potential pitfalls associated with each metric, along with detailed experiments and results.