EC (Edge Computing) is about moving part of the service-specific processing and data storage from the Cloud Computing to the edge network nodes. Among the expected benefits of EC deployment in 5G, there are: performance improvements, traffic optimization and new ultra-low-latency services. If today EC is getting momentum, we’re witnessing, at the same time, a growing development of AI (Artificial Intelligence) for a wide spectrum of applications, such as: intelligent personal assistants, video/audio surveillance, smart cities’ applications, self-driving, Industry 4.0. The requirements of these applications seem calling an AI’s resources-hungry model, whose cloud-centric execution appears in the opposite direction with a migration of computing, storage and networking resources at the edge. In reality, the two technology trends are crossing in the EI (Edge intelligence): an emerging paradigm meeting the challenging requirements of future pervasive services scenarios where optical-radio networks requires automatic real-time joint optimization of heterogeneous computation, communication, and memory/cache resources and high dimensional fast configurations (e.g., selecting and combining optimum network functions and inference techniques). Moreover, the nexus of EI with distributed ledger technologies will enable new collaborative ecosystems which can include, but are not limited to: network operators, platform providers, AI technology/software providers and Users.
If AI is booming and Edge Computing is coming soon
In the last few years, we have witnessed a growing market interest and exploitation of AI solutions in a wide spectrum of ICT applications. AI services are becoming more and more popular, such as intelligent personal assistants, video/audio surveillance, smart cities’ operations, self-driving vehicles, Industry 4.0 . AI flourished again (after the AI winter) due to the current availability of powerful and low-cost processing and storage resources in the Cloud Computing: an abundance meeting the resources hungry requirements of AI methods and systems, called to elaborate un enormous quantity of Big Data.
On the other hand, it is well-known that most of these big data elaborated by AI are generated at the edge, nearer the Users. The wide diffusion of smart terminals, devices and mobile computing, the IoT (Internet-of-Things) and the proliferation of sensors and video cam are generating several hundreds of ZB at the network edge. For example, Cisco estimates that nearly 850 ZB will be generated by people, machines, and things at the network edge by 2021 .
Indeed, the edge is where the Cloud Computing model is also being extended. EC, in fact, is about moving part of the service-specific processing and data storage from the Cloud Computing to the edge network nodes. Among the expected benefits of EC deployment in 5G, there are: performance improvements, traffic optimization and new ultra-low-latency services.
These trends are, obviously, encouraging to consider how extending AI frontier towards the edge ecosystems, in the last mile(s): if fact, the interplay of AI and EC will bring clear advantages in several dimensions. Figure 1 provides a brief summary of some of the cross benefits provided by entangling AI with EC .
Because of these trends, a new emerging innovation avenue is emerging rapidly, called “Edge Intelligence” (EI). This avenue brings several technical challenges: to make it simple, while EC aims at distributing computational and storage resources across the network architecture, real decentralized AI solutions are still in infancy.
The diffusion of EI doesn’t mean, obviously, that there won’t be a future for a centralized CI (Cloud Intelligence). The orchestrated use of Edge and Cloud virtual resources, in fact, is expected to create a continuum of intelligent capabilities and functions across all the Cloudifed infrastructures. This is one of the major challenges for a successful deployment of an effective and future-proof 5G.
How a pervasive AI will bring to new network models
Software Defined Network (SDN) and Network Function Virtualization (NFV) technologies and models are offering the unprecedent opportunity of designing and operating networks infrastructures (e.g., 5G), with two decoupled architectural levels, hardware and software. This is brining higher flexibility and programmability, but, at the end of the day, it doesn’t change the foundations about the network layering and stack protocols, as historically designed with the OSI models and the Internet. It's layers all the way down, where each layer controls a subset of the decision variables, and observes a subset of constant parameters and the variables from other layers. When transmitting data from one terminal to another, the data flow starts at the Application layer proceeding down to the Physical layer; then over the channel to the receiving terminal and back up the hierarchy.
Future services scenarios may take benefits from enhancing this three-decades old model by exploiting a pervasive intelligence. For example, network performance would be radically improved if instead network layering and stack protocols would be systematically designed as distributed solutions to decision making and optimization context-specific problems. In some cases, it might not be necessary entirely crossing up and down the OSI hierarchy.
The idea of a new mathematical theory of networks has been nicely described in , where it is argued that if network dynamics is modeled by a generalized network utility maximization problem, each layer corresponds to a decomposed subproblem, and the interfaces among layers are quantified as functions of the optimization variables coordinating the subproblems. The reverse-engineering of network layering and stack protocols - originally designed with ad hoc heuristics - discovers the original underlying mathematical (decision making and optimization) problems  to be solved. This consequent flexibility would bring to several alternative decompositions, leading to a wider choice of different layering and software architectures. In turn, this would mean greater modularity, flexibility, adaptability and cost effectiveness of the overall network infrastructure.
This is food for an extended AI, which is already dealing with very complex problems requiring classification, regression, optimization and decision making. Embedding pervasive DL (Deep Learning) methods and systems into the network domain it likely to bring to radically new paradigms in network design and management as well as new service models.
As a matter of fact, recent advances in AI might have already considerable impact in different sub¬fields of ongoing networking. The paper  provide an interesting overview of the state-of-art in this direction, together with a number of significative examples. For instance, traffic prediction and classification are two of the most studied applications of AI in the networking field. DL is also offering promising solution for efficient resource management and network adaption thus improving, even today, network system performance (e.g., traffic scheduling, routing and TCP congestion control). Another area where EI could bring performance advantages is efficient resource management and network adaption. Example issues to address include traffic scheduling, routing, and TCP congestion control.
On the other hand, today it is rather challenging to design a real-time system with heavy computation loads and big data. This where EC enters the scene. An orchestrated execution of AI methods in the computing resources not only in the cloud but also at the edge, where most data are generated, will help in this direction. Moreover, collecting and filtering a large amount of data that contain both network profiles and performance metrics is still very critical, and that the question becomes even more expensive when considering the need of data labelling. Even these bottlenecks could be faced enabling EI ecosystems capable of engaging win-win collaborations between Network/Service Providers, OTTs, Technology Providers, Integrators and Users.
A further dimension is that a network embedded pervasive intelligence (Cloud Computing integrated with Edge Intelligence in the network nodes and smarter-and-smarter terminals) could also pave the way to leveraging the achievements of the emerging distributed ledger technologies and platforms (read more on the side).
Towards a network that design and operate itself
The idea that future networks should be able to mitigate a growing complexity by self-organizing is well known: think about the past advances in Autonomic Networking, SON (Self-Organizing Networks). Zero-touch Networks and the adoption of AI methods and systems in OSS/BSS and orchestrations processes.
This trend, today, is accelerating with the idea that AI could (almost) autonomously design, shape and operate the architectures of future networks, and it could do that optimally meeting the techno-economic and business real-time requirements.
It is not new, in fact, that one of the most challenging open question in AI is finding the way to automatically write code from kinds of specifications that humans can easily provide, such as via natural language instruction. A key idea is that a flexible combination of pattern recognition and explicit reasoning can be used to approach that open question. For instance, SketchAdapt , trained on tens of thousands of programs examples, learns how to compose short, high-level programs, while letting a second set of algorithms find the right sub-programs to fill in the details. This is just a very simple example, but it is showing an important trend, growing exponentially. AI might have the potentiality to design and create new network modules and com-ponents with limited or even without human involvement. As a matter of fact, Generative Adversarial Networks (GANs) have also shown that the DL models have these abilities in generating new elements, modules and developing strategies that humans fail to discover with the same rapidity (e.g., AlphaGo).
A reverse side of the coin: AI is still power hungry and time consuming
The technical reason why DNNs (Deep Neural Networks) are so performing is not fully understood yet. One possible explanation, described in literature, is that being DNNs based on an iterative coarse-graining scheme, their functioning is somehow rooted to some fundamental theoretical physics laws or tools (e.g., Renormalization Group).
It might not be a surprise that DNNs functioning resembles some profound law of Nature (imagine the complex neurons networks in the brain); but in one thing we are still very far away from emulating Nature: energy consumption.
Current AI solutions are quite resources/energy-hungry and still time-consuming. In fact, today DNNs (as other AI models) still rely on Boolean algebra transistors to do an enormous amount of digital computations over huge data sets. The roadblock is that chipsets technologies aren’t getting faster at the same pace as AI software solutions are progressing in serving markets’ needs. Today, for example, Cloud servers and data centers currently account for around 2% of power consumption in the U.S. According to some forecasts, data centers will consume one fifth of the world’s electricity by 2025 .
Will this energy consuming trend be really sustainable in the long term? We remind that in basic functioning of a DNN, each high-level layer learns increasingly abstract higher-level features, providing a useful, and at times reduced, representation of the features to a lower-level layer. This similarity is suggesting , the intriguing possibility that DNNs principles are deeply rooted in Quantum Field Theory and Quantum Electromagnetics . This aspect is, perhaps, offering a way to bypass above roadblocks: developing AI technologies based on photonic/optical computing systems which are much faster and much less energy-consuming that current ones.
As a matter of fact, while, in line with the Moore’s law, electronics starts facing physically fundamental bottlenecks, nanophotonics technologies  are considered promising candidates to overcome electronics future limitations. Consider that DNNs operations are mostly matrix multiplications, and nanophotonic circuits can make such operations almost at the speed of light and very efficiently due to the nature of photons. In simple words, photonic/optical computing uses electromagnetic signals (e.g., via laser beams) to store, transfer and process information. Optics has been around for decades, but until now, it has been mostly limited to laser transmission over optical fiber. Today technologies, using optical signals to do computations and store data, would accelerate AI computing by orders of magnitude in latency, throughput and power efficiency.
Matrix multiplication, for example, is one of the most expensive and time-consuming calculations involved in DNN, and being able to speed it up will help create much faster DNN models. Optical computing allows making any matrix multiplication, regardless of the size, in one CPU clock, while electronic chips take at least a few hundred clocks to perform the same operations. Research and innovation activities are already producing concrete results and prototypes (Box 2).
Markets movements and start-ups are confirming this trend. Lightellingence , for example, is a start-up proposing a new architecture for a fully optical neural network that, in principle, could offer an enhancement in computational speed and power efficiency over state-of-the-art electronics for conventional inference tasks. LightOn  is another example of start-up building Optical Processing Units (OPUs) which uses light to perform computations matrix-vector multiplications with non-linearities for AI. The OPU can just be plugged onto a standard server or workstation, and accessed through a toolbox that ca be integrated within familiar programming environments.
A next exploitation of Edge Computing in 5G will allow migrating part of the service-specific processing and data storage from the Cloud Computing to the edge network nodes. Expected benefits includes: performance improvements, traffic optimization and new ultra-low-latency services.
Today, in this scenario, we are also witnessing a growing development of Artificial Intelligence solutions for a wide spectrum of applications and services. The evolving requirements are demanding an AI more and more
resources-hungry model. Therefore cloud-centric execution appears in the opposite direction with a migration of computing, storage and networking resources at the edge. In reality, the intertwining of Edge Computing and Artificial Intelligence trajectories has already given rise to a new emerging innovation trend, called Edge Intelligence.
The orchestrated use of Edge and Cloud virtual resources, in fact, is expected to create a continuum of intelligent capabilities and functions across all the Cloudifed infrastructures. This is one of the major challenges for a successful deployment of an effective and future-proof 5G.
A major roadblock to this vision is the long-term extrapolations of the energy consumption needs of a pervasive Artificial Intelligence embedded into future network infrastructures. Low-latency and low-energy neural network computations can be a game changer. In this direction, fully optical neural network could offer impressive enhancements in computational speed and reduced power consumptions.
 Zhou, Zhi, et al. "Edge Intelligence: Paving the Last Mile of Artificial Intelligence with Edge Computing." arXiv preprint arXiv:1905.10083 (2019)
 Cisco Global Cloud Index: Forecast and Methodology, 20162021 White Paper. [Online]. Available: https://www.cisco.com/c/en/us/solutions/collateral/service-provider/global-cloud-index-gci/white-paper-c11-738085.html
 Lovén, Lauri, et al. "EdgeAI: A Vision for Distributed, Edge-native Artificial Intelligence in Future 6G Networks." The 1st 6G Wireless Summit (2019): 1-2
 Chiang, Mung, et al. "Layering as optimization decomposition: A mathematical theory of network architectures." Proceedings of the IEEE 95.1 (2007): 255-312
 Wang, Mowei, et al. "Machine learning for networking: Workflow, advances and opportunities." IEEE Network 32.2 (2017): 92-99.
 Nye, Maxwell, et al. "Learning to infer program sketches." arXiv preprint arXiv:1902.06349 (2019).
 Lee, J.W. Quantum fields as deep learning. arXiv preprint arXiv:1708.07408 2017
 Manzalini, Antonio. "Complex Deep Learning with Quantum Optics." Quantum Reports 1.1 (2019): 107-118
 Yao, K.; Unni, R.; Zheng, Y. Intelligent nanophotonics: merging photonics and artificial intelligence at the nanoscale. arXiv preprint arXiv:1810.11709 2018