Transposed convolutional neural networks for generating synthetic waveforms

We will introduce a transposed convolutional neural network that accelerates our pulsar waveform modeling, resulting in a >300x speedup compared to traditional methods. Notably, our neural network architecture includes no structures inherently designed for pulsar modeling or any specific physical parameter details, such that our network architecture is applicable to various waveform or spectrum modeling tasks. We will discuss our neural network design decisions with the intent of providing insights into how other high-energy researchers can introduce neural network methods into their own work. This will include a brief intuitive understanding of which network structures, both high- and low-level, are applicable to this domain and why some, though commonly used, are detrimental for these purposes. We will also cover some of the pitfalls and limitations that may be encountered and how to mitigate them. Finally, we will demonstrate how a neural network trained for one waveform generation task can be retrained for another waveform or spectrum generation task while minimizing the required training data.