information-bottleneck-theory
[ICMLW 2023, IEEE CISS 2024] Assessing Neural Network Representations During Training Using Noise-Resilient Diffusion Spectral Entropy