Optimizing training of time series diffusion models via similarity score functions: application to cyclic and acyclic...

Denoising diffusion probabilistic models have shown the capability to generate synthetic sensor signals. These models rely on a loss function that measures the difference between the noise added during the forward process and the noise predicted by the diffusion model, thereby enabling realistic data generation. However, the stochastic nature of the process and the loss function complicate the estimation of data quality.

To address this issue, we evaluated multiple similarity metrics and adapted an existing metric to monitor both the training and data synthesis processes. The adapted metric was further fine-tuned on the input data to align with the requirements of a downstream classification task.

By incorporating the adapted metric, we significantly reduced the number of training epochs required without observing performance degradation in the classification task.

Our findings demonstrate that optimizing the training process using similarity metrics not only conserves computational resources but also shortens the training time for generative models, making them more efficient and practical for real-world applications.