Speaker
Description
We present results of generative deep learning as applied to a large database of solar images and discuss challenges in training and validation, in particular with distributed training on GPU supercomputers. Our dataset is based on SDO EUV data with high resolution and with excellent coverage since 2010. This dataset is thus well suited to study the application of advanced machine learning techniques that require large amounts of data for training, such as deep learning approaches. We focus on diffusion type models and generative adversarial deep learning (GANs). We address the potential of data augmentation techniques for improved learning and image quality and the opportunities for latent space structure exploration and control. The challenges in reaching the highest resolutions with good image fidelity are analyzed. Potential application downstream that can make use of such generated images are briefly discussed and the need for a community-driven, physics-based basis to establish evaluation criteria for generative models will be emphasized.
| Submit to 'solar physics' topical issue? | Maybe |
|---|
