AI for Robust Compression

Data nowadays is created at a staggeringly pace. Popular websites displaying user-generated content, for example, can generate more than 4.5M pictures and more than 70,000 hours of video in a single day. This poses diverse challenges in terms of storage, transmission, and retrieval of large quantities of such multimedia data. Although the last decades have witnessed great advances in compression algorithms, these rely on universal principles and, therefore, do not easily adapt to data. In the last years, however, major breakthroughs in compression rates have been achieved with deep neural networks which, via their ability to extract underlying structures from large quantities of data, can easily adapt to the data they are fed. While this allows state-of-the-art compression rates, it also makes the compression algorithms less robust to interference.

The goal of this PhD project is to create algorithms that are both compression efficient and robust to interference. This will be done by exploring how state-of-the-art neural network architectures can be combined with recent sparse inference algorithms, and also with classic compression techniques.