Optimization for Networks and Object Recognition

The present thesis explores two different application areas of combinatorial optimization, the work presented, indeed, is two fold, since it deals with two distinct problems, one related to data transfer in networks and the other to object recognition. Caching is an essential technique to improve throughput and latency in a vast variety of applications. The core idea is to duplicate content in memories distributed across the network, which can then be exploited to deliver requested content with less congestion and delay. In particular, it has been shown that the use of caching together with smart offloading strategies in a RAN composed of evolved NodeBs (eNBs), AP (e.g., WiFi), and UEs, can significantly reduce the backhaul traffic and service latency. The traditional role of cache memories is to deliver the maximal amount of requested content locally rather than from a remote server. While this approach is optimal for single-cache systems, it has recently been shown to be, in general, significantly suboptimal for systems with multiple caches (i.e., cache networks) since it allows only additive caching gain, while instead, cache memories should be used to enable a multiplicative caching gain. Recent studies have shown that storing different portions of the content across the wireless network caches and capitalizing on the spatial reuse of device-to-device (D2D) communications, or exploiting globally cached information in order to multicast coded messages simultaneously useful to a large number of users, enables a global caching gain. We focus on the case of a single server (e.g., a base station) and multiple users, each of which caches segments of files in a finite library. Each user requests one (whole) file in the library and the server sends a common coded multicast message to satisfy all users at once. The problem consists of finding the smallest possible codeword length to satisfy such requests. To solve this problem we present two achievable caching and coded delivery scheme, and one correlation-aware caching scheme, each of them is based on a heuristic polynomial-time coloring algorithm. Automatic object recognition has become, over the last decades, a central toping the in the artificial intelligence research, with a a significant burt over the last new year with the advent of the deep learning paradigm. In this context, the objective of the work discussed in the last two chapter of this thesis is an attempt at improving the performance of a natural images classifier introducing in the loop knowledge coming from the real world, expressed in terms of probability of a set of spatial relations between the objects in the images. In different words, the framework presented in this work aims at integrating the output of standard classifiers on different image parts with some domain knowledge, encoded in a probabilistic ontology.