Over the past few years, deep learning models have been increasingly applied to domains where data is naturally represented with complex numbers, such as MRI data and radar signals. In such scenarios using a real valued neural network fails to fully exploit the correlation between the real part and the imaginary part.
This poster presents the use of complex-valued convolutional neural networks (CV-CNNs) for a radar signal processing case study aided by packages such as torch-complex and complexPytorch. The poster will include: (1) Illustrative visuals showing how complex convolutions operate and how they preserve phase relationships, along with side-by-side comparisons highlighting the superior performance of complex-valued networks over real-valued networks on complex signal data; (2)Illustrations of workflows for generating training datasets with practical radar effects such as array imperfections, inhomogeneous clutter and low SNR scenarios; (3) Through illustrative figures, you can see how to build complex valued neural networks using tools from complexPytorch. This section emphasizes the extensibility of Python ecosystem, showing how packages such as torch- complex and complexPyTorch can be adapted for a variety of scientific applications.
Whether you’re coming from a deep learning, signal processing, or applied research background, this poster will provide practical insights and tools for working with complex-valued data pipelines in Python, and showcase how PyTorch can be extended to support emerging complex-valued deep learning applications.