Complex Deep Networks
Dear students, if you are seeking a topic for a master thesis, we currently offer the following subject supervised by my PhD student Fatima Sajid Butt:
Complex valued Neural Networks:
Complex numbers are everywhere around us. From electrical circuits to signal processing, they have a huge application spectrum. They constitute essential part of many mathematical models.
Complex numbers are not represented very often in deep learning as complex numbers but mostly modelled as real numbers for sake of simplicity. Though convenient, it does not always serve the purpose. Specifically when complex numbers have to be represented in their true forms to achieve accurate results. A complex valued Neural network is an extension of a real-valued neural network, whose input and output signals, parameters (weights) are all complex numbers (hence the activation function is a complexed-valued function). Many machine learning platforms such as Matlab do not offer any explicit support to deal with complex numbers in neural networks.
This thesis focuses on developing a toolbox to deal with complex numbers and their representation in deep neural networks e.g. in LSTMs(Long Short Term Memory) or CNNs (convolutional neural networks). TensorFlow, an end-to-end opensource machine learning platform is recommended. A strong background in C++ or/and Python is essential.