Tag
This paper proposes HyMLRaman, a hybrid framework combining EfficientNet-B3 spectral embeddings with classical classifiers (notably SVM) and DDPM-based generative feature augmentation to identify six pharmaceutical compounds via Raman spectroscopy, achieving 96.31% accuracy.
This paper provides a theoretical analysis explaining why deterministic DDIM samplers hallucinate more than stochastic DDPM samplers in diffusion models, attributing it to getting stuck in mode-interpolation regions during reverse dynamics.