ABSTRACT
Generative models have burst onto the world stage, in the form of large language models and image generation tools. A generative model aims to capture the statistical distribution of a dataset, in order to be able to generate new datapoints by sampling from this distribution. This chapter covers some history leading up to the advent of generative deep learning models, discussing how the problem of task specificity is overcome by transfer learning and fine-tuning. Large language models based on the transformer neural network architecture are described, together with some of their current limitations of hallucination, bias, training data memorization, and collapse in output diversity. Some social and legal implications of generative AI are discussed, both for text and image generation techniques. Applications of deep learning architectures and techniques to the problem of numerical weather prediction are then described, which recently overtook classical weather models in predicting geopotential, temperature, specific humidity, and wind speed at 500 hPa, for certain lead times. Generative models applied to remote sensing are also briefly discussed, and finally, some possible future developments of these novel approaches are considered.
