Conditioned Cycles in Sparse Data Domains: Applications to Electromagnetics

Maria Barger, Randy C. Paffenroth, Harsh Nilesh Pathak

Abstract: Training neural networks is difficult with small datasets, yet small training data sets are commonly found with machine learning problems in the physical sciences. Prediction accuracy within such sparse data domains can be improved by the utilization of conditioning, inspired by classical statistics, and cycles, similar to those of CycleGANs. In this paper, we propose the addition of “oracle“ conditioning, where the true condition value is provided during training (but not during testing), to improve final prediction accuracy in sparse data domains. For example, in over-water electromagnetic (EM) source localization, the range of the emitter can be thought of as a “condition“ on the generation of a received EM signal. Oracle conditioning utilizes an encoder-decoder framework where the conditioning of the decoder is not contingent on the output of an encoder's prediction (e.g, predicting the range from a received EM signal). While our primary application involves source localization for EM signals, this paper is intended to inspire a general approach toward neural network training for problems in the physical sciences with similar characteristics. Our multi-component loss function is intended to serve as a general method for small data problems which can benefit from the addition of cycles and oracle conditioning.