Eduardo Cueto Mendoza

Title: Model Compression for Deep Neural Networks

Supervision Team: John Kelleher, TUD / Rozenn Dahyot, Maynooth University

Description: Deep learning has revolutionized digital media analysis, be it video video, image, text, or indeed multimedia. The deep learning revolution is driven by three long-term trends: Big Data, more powerful computers (GPUs/TPUs), and ever larger models. At the same time there has been an increase in edge computing and the deployment of deep neural networks to devices that have limited resources (such as memory, energy or bandwidth). This project will explore the development of novel cost functions, tailored to deep learning models for video and image analysis; compression techniques for deep neural networks for video and image analysis; and error analysis for model compression techniques.