Content area
This thesis uses machine learning to help develop an understanding of how the physics and geometry of different underwater objects affect how SONAR interacts with them. The key contributions are designing novel machine learning tools that use complex numbers, which incorporate phase information critical to physical interpretation of sonar targets.
The primary machine learning approach used is an autoencoder--a data compression neural network that reduces the size of data vectors. This work evaluates different levels of autoencoder compression for data containing SONAR interactions with various targets. Some physical properties of the targets and experimental environments can be autonomously extracted from the compressed data such as the distance between a target and SONAR device, which is information of high naval interest.
Another key contribution of this work is the introduction of new types of neural network layers called discrete transform layers. They extend a well-studied technique in the field of signal processing called the Fourier Transform. This thesis demonstrates that the novel discrete transform layers improve the learning power of neural networks that use them instead of the classical Fourier Transforms. Their unique properties enable a novel processing block that outperforms the closest equivalent standard machine learning layer, a convolutional layer, in single-channel contexts.