Continuous representation of kernels used in convolutional neural networks

Within the field of machine learning the research of Convolutional Neural Networks (CNN) has been rapidly progressing. A recent development has been representing the kernels used in the convolutional layers continuously through a separate auxiliary neural network. A practical application of a CNN is in medical imaging, where the neural network is trained to detect micro-bubbles in order to map the vascular network. This work investigates the use of continuous kernel representations for the problem of micro-bubble localization. This is done by training the neural network on a simulated ultrasound signal where the result aims to replicate the corresponding ground truth bubble location.