Adaptive Latent Space Tuning for Non-Stationary Distributions

Encoder-decoder style deep convolutional neural networks (CNN) are able to extract features directly from images, mix them with scalar inputs within a general low-dimensional latent space, and generate outputs which represent complex physical phenomenon. One challenge faced by deep learning methods is modeling large non-stationary systems whose characteristics change quickly with time for which re-training is not feasible. In this paper we present a method for adaptive tuning of the low-dimensional latent space of deep encoder-decoder style CNNs based on real-time feedback to compensate for unknown and fast distribution shifts. We demonstrate the approach for predicting the properties of a time-varying charged particle beam in a particle accelerator whose initial distribution and components (accelerating electric fields and focusing magnetic fields) are quickly changing with time and may not be measurable during operations. Our method utilizes the low-dimensional latent space basis directly to generate new outputs and therefore does not require access to new input beam distributions for re-training, which is important for large systems such as particle accelerators where input distribution measurements interrupt normal operations.