Skip to main content

Context- and Task-Oriented Communication

Periodic Reporting for period 2 - CTO Com (Context- and Task-Oriented Communication)

Reporting period: 2018-11-01 to 2020-04-30

New infrastructure in 5G networks, such as cloud-RANs, caches, relay-drones, and massive MIMO arrays, bring enormous improvements in data rates and delays even when applied naively using standard coding techniques. However, to satisfy future demands, new communication techniques perfectly tailored to exploit such infrastructure will be essential. In fact, the communication scenarios induced by these new elements differ from traditional scenarios significantly with respect to the terminals' contexts. For example, in networks with caches (i.e. additional storage spaces), terminals can download parts of their desired data directly from closeby caches. On the other hand, relay-drones can reinforce a transmit signal so as to provide a terminal with a stronger second observation of a desired signal. Finally, cloud-RANs allow terminals to obtain information about other transmit and receive signals. There is thus a basic need for highly improved context-oriented communication techniques. This need is even
more immediate in view of the rapidly increasing number of distributed decision and control systems, where the various nodes measure highly correlated signals, and thus have a priori side-information about other nodes' signals. A second major difference between communication in distributed decision and control systems and traditional communication scenarios is their task. In these systems the final task is no more to convey and reconstruct sequences of data bits or observed signals, but to make distributed decisions or take distributed actions that attain a common goal. The traditional approach uses standard LDPC, Polar, or Turbo codes to exchange scalar- or vector-quantizations of the signals measured at the various nodes, and then runs decision, control, or prediction algorithms locally based on all the accumulated information. This approach can be highly suboptimal. In particular for situations where the decisions take value in a small range, the approach can lead to huge amounts of unnecessarily transmitted data. With new, task-oriented communication techniques we aim to drastically reduce this overhead.
During the first period, the project concentrated on finding communication techniques and fundamental limits of cache-aided systems, distributed decision systems, and distributed big data applications. For cache-aided systems we proposed secure piggyback coding, a technique that can communicate close to the fundamental limits of the system while keeping the transmitted data secure from an external eavesdropper. We also proposed new optimal coding techniques for relay combination networks with cache memories. In our study of distributed decision systems, we focused on multi-hop networks, noisy channels, and on cooperative scenarios. We proposed new coding techniques based on unequal error protection mechanisms and unanimous decision forwarding strategies for such distributed decision systems. Our new techniques improve over the previous techniques and in some cases can be proved to attain the optimal error decays. For distributed big data applications, we focused on MAP-Reduce systems and proposed a new framework to describe and solve such systems. We considered systems without and with straggling nodes, for which we proposed simple schemes that are amenable for practical implementation, and we showed that they perform close to the fundamental performance limits. Additionally, we also presented improved communication techniques for multi-antenna wireless optical communication systems and for networks with mixed-delay constraints. To summarize, we presented various new specifically-tailored communication techniques for applications with non-traditional contexts or tasks, and we derived information-theoretic bounds on the fundamental limits of such applications to show that our techniques are almost optimal.
We derived the fundamental performance limits of different distributed hypothesis testing systems. Our focus was on multi-hop and noisy networks, which are particularly relevant for Internet of Things applications were batteries of sensors are meant to last for decades and the transmissions of the sensors thus need to be of low energy. In particular, we showed when it is optimal for the sensors to only relay the decisions at previous sensors but to ignore all the other information transmitted by these sensors. Furthermore, we illustrated the importance of unequal error protection codes for communication in sensor networks. In the remainder of this project we will further consider systems with more stringent energy and delay constraints as well as investigate systems where sensors can sample and communicate sequentially and adaptively.

In another line of work, we presented and analyzed new secure and non-secure cache-aided communication techniques based on our piggyback coding idea. We further showed that in heterogeneous networks, it is beneficial to assign cache memories in an asymmetric way, according to the strengths of the channels. In the remainder of this project we will implement our coding techniques on a platform and hopefully show their relevance in improving practical cache-aided communication systems.

In yet another line of work we designed new coding techniques that can accommodate different information flows, some of them subject to very limited delay constraints. We further prove that the coding schemes attain the fundamental performance limits in certain scenarios. Such mixed-delay networks are particularly important also for control applications or coordination of smart agents because these applications typically have very stringent delay constraints but coexist with applications that are less sensitive to delays. In futures work we will concentrate both on communication for control and on communication for coordination purposes.

Our forth line of work established optimal coding and computing schemes for the popular MAP-Reduce framework. In particular, we treated scenarios with unreliable computing nodes, where any node can fail with a certain probability. In the remainder of this project we will consider other big data applications such as clustering or analysis of deep neural networks.

Our last line of work of ours considers new coding techniques for multi-antenna free-space (visible light or infra-red) communication where we proposed a new antenna cooperation strategy and showed that it is optimal at high and low signal-to-noise ratio (SNR), and it achieves improved performances also in the moderate SNR regime.

To summarize, in this project we made significant progress with new context- and task-oriented communication techniques for various applications. In the rest of the project we will further investigate these scenarios but also branch out to other practically relevant setups.