A neural network  is a mathematical object inspired by the functioning of the neurological systems of animals. A complex (or deep neural network, from which Deep Learning by extension) is a neural network with many computing layers, much more than traditional neural networks. A computing layer applies simple mathematical operations to its inputs (typically real numbers vectors) that produce new numeric vectors as result , which in turn enter the subsequent layers of the network. There are many types of layers in the scientific literature. In the field of image classification, the layers that have proved to be fundamental, contributing performance that clearly advances the state of the art provided by traditional technologies, are the so-called “convolutional” layers (see figure).

Another distinctive feature of Deep Learning technologies is the nature of their input vectors, that is, the numeric carriers at the entrance of the first layer. Unlike the traditional Machine Learning methods – which normally use numeric vectors that represent some of the content features that are considered important for the specific task – input to complex networks, in the case of image classification, are the images themselves.

In this extremely vivid and continuous context, the project aims to study new deep network methods and architectures that are applicable to the domain of video objects, which today represent a real challenge for these technologies due to their high dimensionality and semantic complexity.

The realization of these architectures would make the RAI archive metadata process more efficient while at the same time increasing the quality and richness of content-related information. The types of information that can be extracted effectively through these networks are not limited to generic classification, although the latter can be achieved principally using an arbitrary classification system (content-oriented, emotional, and recognizing faces of known characters). More articulated versions can extract structured annotations, substantially reducing the gap with traditional documentation systems.