分野別セミナー

【講演要旨】It has been a fundamental challenge in deep learning research to understand the internal dynamics of neural networks. To address this challenge, in this study we apply the Mapper algorithm, a topological data analysis (TDA) method, to visualize and analyze the weight evolution of feedforward neural networks (FNNs) during training. We construct Mapper graphs to capture the changes in the weight space. Our experiments on the MNIST dataset with varying hidden layer sizes reveal that these topological representations effectively highlight key training phases, including class separability and misclassification. The results demonstrate that Mapper-based visualizations provide valuable insights into how weight distributions evolve over time, offering a complementary perspective to conventional evaluation metrics. This approach contributes to a better qualitative understanding of neural network learning dynamics.