Distinguished Seminar Series on Data Science & Artificial Intelligence - Advanced Machine Learning and Applications by Prof. Moncef Gabbouj

Rethinking Deep Learning (DL) by reconsidering the core neuron model used in all ANN architectures for image processing applications. DL outperformed many traditional approaches in numerous fields of science. However, DL comes at a price of high computational cost and follows mostly a Blackbox approach. Targeting Green Learning, we aim to develop more computationally efficient Artificial Neural Networks, called Operational Neural Networks as alternatives to conventional Convolutional Neural Networks (CNNs). ONNs can perform any linear or non-linear transformation with a proper combination of “nodal” and “pool” operators. This is a great leap towards expanding the neuron’s learning capacity in CNNs, which thus far required the use of a single nodal operator for all synaptic connections for each neuron. This restriction has been lifted by introducing a superior neuron called the “generative neuron” where each nodal operator can be customized during the training to maximize learning. As a result, the network can self-organize the nodal operators of its neurons’ connections. Self-Organized ONNs (Self-ONNs) equipped with superior generative neurons can achieve diversity even with a compact configuration. We shall explore the use of neural network models equipped with the generative and the super neuron in several image and video processing applications.