#23: MLDublin meets ADAPT @ DogPatch Labs- 1 min
Replacing the background behind an actor with the use of a greenscreen is a common task in the film and television industry. However, in the rise of amateur user generated content on YouTube, and video messaging the greenscreen proves to be a undesirable and impractical constraint. By leveraging deep learning we are able to separate foreground and background elements in a natural environment without large user interaction thanks to semantic understanding of a scene. In this talk I give a brief overview of greenscreen and natural image keying and it's relation the field of semantic segmentation. I will discuss in particular recent work of ours, submitted to ICIP, where we experiment with the benefits of a multi-task objective for the task of natural image keying and the possibilities we open for future work.
Machine learning has emerged as a new clever way for optimizing and improving the performance of fiber-optic telecommunication systems by tackling both deterministic and stochastic noises in the network without increasing complexity. The potential of developing new modems incorporating machine learning technology to provide consistently high-speed broadband connectivity is an exciting new research topic. Digital signal processing (DSP)-based machine learning bridges the gap between all-optical high-resolution signal processing and DSP. Harnessing appropriate machine learning algorithms we can successfully compensate nonlinear effects in electronic domain to increase transmission-reach of modern high-capacity optical systems.