Lec Assignment4 2020

This case study covers progress in the combined area of optics and AI, ranging from
photonic neural network (NN) architecture to AI-enabled advances in optical communications
including both physical layer transceiver signal processing and network layer performance
monitoring,
1. Photonic neural network architecture
2. AI for short-reach optical communications
3. AI for medium- and long-reach optical communications
4. AI for optical performance monitoring
The progress in the field of PNN is indeed remarkable,
5. especially in the latest years, and new promising research
6. directions are emerging, with potentially disruptive impact
7. on developing faster and/or less energy-hungry PNN imple-
8. mentations. Many challenges are still to be overcome: many
9. diverse material platforms are being investigated without a
10. clear winner, the maturity of the developed hardware must
11. improve (in many demonstrators critical elements are still
12. emulated in electronics), the scalability and the robustness of
13. the proposed PNN architectures should grow in order to meet
14. the requirements of typical machine learning problems.
15. Nevertheless the growing research efforts both from
16. academia, prominent companies, and fast-growing start-ups
17. indicate that the coming years will witness a huge expansion
18. of this field, with strong potentials to address real-world
19. applications.
Photonic neural network architecture
The progress in the field of PNN is indeed remarkable,
especially in the latest years, and new promising research
directions are emerging, with potentially disruptive impact
on developing faster and/or less energy-hungry PNN imple-
mentations. Many challenges are still to be overcome: many
diverse material platforms are being investigated without a
clear winner, the maturity of the developed hardware must
improve (in many demonstrators critical elements are still
emulated in electronics), the scalability and the robustness of
the proposed PNN architectures should grow in order to meet
the requirements of typical machine learning problems.
Nevertheless the growing research efforts both from
academia, prominent companies, and fast-growing start-ups
indicate that the coming years will witness a huge expansion
of this field, with strong potentials to address real-world

JASMINE[17105051] Page 1
 Lec Assignment4 2020

applications.
The progress in the field of PNN is indeed remarkable,
especially in the latest years, and new promising research
directions are emerging, with potentially disruptive impact
on developing faster and/or less energy-hungry PNN imple-
mentations. Many challenges are still to be overcome: many
diverse material platforms are being investigated without a
clear winner, the maturity of the developed hardware must
improve (in many demonstrators critical elements are still
emulated in electronics), the scalability and the robustness of
the proposed PNN architectures should grow in order to meet
the requirements of typical machine learning problems.
Nevertheless the growing research efforts both from
academia, prominent companies, and fast-growing start-ups
indicate that the coming years will witness a huge expansion
of this field, with strong potentials to address real-world
applications.
The progress in the field of PNN is indeed remarkable,
especially in the latest years, and new promising research
directions are emerging, with potentially disruptive impact
on developing faster and/or less energy-hungry PNN imple-
mentations. Many challenges are still to be overcome: many
diverse material platforms are being investigated without a
clear winner, the maturity of the developed hardware must
improve (in many demonstrators critical elements are still
emulated in electronics), the scalability and the robustness of
the proposed PNN architectures should grow in order to meet
the requirements of typical machine learning problems.
Nevertheless the growing research efforts both from
academia, prominent companies, and fast-growing start-ups
indicate that the coming years will witness a huge expansion
of this field, with strong potentials to address real-world
applications.
The progress in the field of PNN is indeed remarkable,
especially in the latest years, and new promising research
directions are emerging, with potentially disruptive impact
on developing faster and/or less energy-hungry PNN imple-
mentations. Many challenges are still to be overcome: many
diverse material platforms are being investigated without a
clear winner, the maturity of the developed hardware must
improve (in many demonstrators critical elements are still

JASMINE[17105051] Page 2
 Lec Assignment4 2020

emulated in electronics), the scalability and the robustness of

the proposed PNN architectures should grow in order to meet
the requirements of typical machine learning problems.
Nevertheless the growing research efforts both from
academia, prominent companies, and fast-growing start-ups
indicate that the coming years will witness a huge expansion
of this field, with strong potentials to address real-world
applications.
The progress in the field of PNN is indeed remarkable,
especially in the latest years, and new promising research
directions are emerging, with potentially disruptive impact
on developing faster and/or less energy-hungry PNN imple-
mentations. Many challenges are still to be overcome: many
diverse material platforms are being investigated without a
clear winner, the maturity of the developed hardware must
improve (in many demonstrators critical elements are still
emulated in electronics), the scalability and the robustness of
the proposed PNN architectures should grow in order to meet
the requirements of typical machine learning problems.
Nevertheless the growing research efforts both from
academia, prominent companies, and fast-growing start-ups
indicate that the coming years will witness a huge expansion
of this field, with strong potentials to address real-world
applications.
The progress in the field of PNN is indeed remarkable,
especially in the latest years, and new promising research
directions are emerging, with potentially disruptive impact
on developing faster and/or less energy-hungry PNN imple-
mentations. Many challenges are still to be overcome: many
diverse material platforms are being investigated without a
clear winner, the maturity of the developed hardware must
improve (in many demonstrators critical elements are still
emulated in electronics), the scalability and the robustness of
the proposed PNN architectures should grow in order to meet
the requirements of typical machine learning problems.
Nevertheless the growing research efforts both from
academia, prominent companies, and fast-growing start-ups
indicate that the coming years will witness a huge expansion
of this field, with strong potentials to address real-world
applications.
The progress in the field of PNN is indeed remarkable,

JASMINE[17105051] Page 3
 Lec Assignment4 2020

especially in the latest years, and new promising research

directions are emerging, with potentially disruptive impact
on developing faster and/or less energy-hungry PNN imple-
mentations. Many challenges are still to be overcome: many
diverse material platforms are being investigated without a
clear winner, the maturity of the developed hardware must
improve (in many demonstrators critical elements are still
emulated in electronics), the scalability and the robustness of
the proposed PNN architectures should grow in order to meet
the requirements of typical machine learning problems.
Nevertheless the growing research efforts both from
academia, prominent companies, and fast-growing start-ups
indicate that the coming years will witness a huge expansion
of this field, with strong potentials to address real-world
applications.
The progress in the field of PNN is indeed remarkable, especially in the latest years, and new
promising research directions are emerging, with potentially disruptive impact on developing
faster PNN implementations. Nowadays Artificial Intelligence is used in an impressively
large number of applications, comprising image classification, speech recognition and
language translation, decision making, web searches, content filtering on social networks,
recommendations on e-commerce websites, etc.
Neural Networks are useful for processing large data sets, combining and analyzing vast
amounts of information quickly and without the need of explicit instructions. For the
implementation of massively interconnected ANN, the conventional computer architecture is
fundamentally inefficient and not scalable with respect to computation, memory, and
communication.
To address the shortcomings of today’s computer architecture for neural networks with the
aim of increasing the computing speed and power efficiency, a growing effort has focused
on the development of specifically tailored electronic architectures. Graphics Processing
Units (GPU) have been identified as particularly suitable for implementing the parallel
computing tasks typical of ANN, and significantly contributed to the current success of
machine learning in real application scenarios.
To this aim, these novel electronic solutions focus on advanced numerical representations,
memory architectures suitable for high-speed matrix multiplications, and a very high
bidirectional off-chip bandwidth (exceeding a Tb/s) to enable model and data parallelism.
With the aim of developing a new integrated circuit technology for ANN where low-voltage
field-effect transistors and non-volatile memories are tightly integrated exploiting quantum
engineering of heterostructures of two-dimensional materials. All the research and
development activities aim to improve both speed and energy efficiency of machine learning
tasks.

JASMINE[17105051] Page 4
 Lec Assignment4 2020
Over the years, photonic solutions for optical communication
and processing evolved along the same lines, aiming at increasing the transmission speed and
the energy efficiency. For this reason, optical implementations of neural networks have been
investigated since a long ago, aimed at exploiting the large parallelism (through degrees of
freedom such as wavelength, polarization, and mode) and the high connectivity achievable
with optics. Additionally, many linear transformations can be performed with passive optics
without power consumption and with minimal latency. These features indicate that optical
implementations of neural networks can overcome electronic solutions in terms of
computational speed and energy efficiency.

Various approaches pursued in the field of Photonic Neural Networks (PNN) – alternatively
called, sometimes, photonic neuromorphic computing. When it comes to incorporating
intelligence to optical systems or networks, three noteworthy topics are :

Handling uncertainty: In an optical network there are various non-deterministic events

taking place and lack of full information about the environment is a common problem.
Therefore, intelligent agents must be able to operate under uncertainty in a robust way. The
laws of probability and, in particular, Bayesian networks are useful tools to build those robust
models:

 Bayesian Networks (BN) offer a unique solution to the next generation of AI usage on

real data since they have the ability to encode causal relationships (for predicting the
future) and to go one step further with its property of “information propagation through
the network.

 BNs provide real-time solutions that are beyond the capabilities of human reasoning
and present them in simple, easy to interpret graphical models

Moreover, optical systems and networks are subject to constant changes. Hence, intelligent
agents must include inference algorithms for temporal models to perform tasks like filtering,
prediction or smoothing, relying on techniques like hidden Markov models (HMM) .
Tackling decision making:
A second key element is the use of decision-making algorithms. The underlying principle for
these algorithms is the maximization of the expected utility, in which a utility function is
defined in order to assign a single number to express the desirability of a state and an agent
makes decisions with the aim of maximizing such a function .
Realistic networking environments, however, must deal with
uncertainty and the utility of an agent usually depends on a sequence of decisions rather than
on a single isolated one. Decision making in optical network agents can therefore be modelled
as sequential decision problems in uncertain environments. These problems can be solved
by Markov decision processes (MDPs) if the agent's actions depend only on the current state
of the agent, and not on its history. MDPs are defined by a transition model, which specifies
the probabilistic outcomes of actions, and by a reward function, which specifies the reward in
each state. The solution of an MDP is a policy that associates a decision with every state that
the agent might reach. An optimal policy maximizes the utility of the state sequences
encountered when it is executed.
Learning:The third issue of paramount importance is learning. Learning enables an agent to
improve its performance on future tasks due to acquired experience. The inclusion of learning
JASMINE[17105051] Page 5
 Lec Assignment4 2020
is important for several reasons. A learning-capable agent can adapt to changes in the
environment and it is even able to adapt to unforeseen scenarios that could not be anticipated
when the agent was designed. Moreover, in many cases, learning from existing data may
be the only way to generate a working model, or in other words, sometimes human
programmers (or engineers) have no idea on how to program a solution themselves.
Statistical learning and artificial intelligence here provide the theory and tools to learn
from existing data, which can be gathered in optical communications systems and
networks to monitoring techniques.

Although agents can handle uncertainty by using the methods of probability and decision
theory, they must learn their probabilistic theories from experience. Thus, Bayesian learning
methods formulate learning as a form of probabilistic inference using the observations to
update a prior distribution over hypotheses; maximum a posteriori (MAP) learning  selects a
single most likely hypothesis given the data, and maximum-likelihood learning simply selects
the hypothesis that maximizes the likelihood of the data. These techniques have been used
in optical receivers.

Apart from the above mentioned techniques machine learning has also been widely used.
There are three main categories in machine learning:
 In supervised learning an agent observes some example input-output pairs and learns a
function that maps from input to output. Techniques include linear regression, logistic
regression, decision trees, artificial neural networks, nearest neighbor models
and support vector machines (SVM) to name just a few. Moreover, different models
can be combined in ensemble learning, with the aim of improving results. Supervised
learning has been used, for instance, for optical performance monitoring, to estimate
the quality of transmission (QoT) in optical networks and for resource allocation in
data centers.
 In unsupervised learning, an agent learns patterns from the input even though no
explicit output is supplied. For instance, clustering and principal component analysis
methods, which belong to this type of learning, have been used for optical performance
monitoring, modulation format recognition and impairment mitigation.
 Finally, in reinforcement learning an agent learns an optimal (or nearly optimal)
policy from a series of reinforcements or punishments received from its interaction
with the environment. Some techniques include adaptive dynamic programming and
temporal-difference (TD) methods. Q-learning , a well-known technique of the latter
type, aims to find an optimal quality value (Q-Value) of action-selection policy for any
given (finite) Markov decision process .For instance, Q-learning has been used for path
and wavelength selection in the context of optical burst-switched (OBS) networks.

JASMINE[17105051] Page 6
 Lec Assignment4 2020
Photonics implementation of AI and neuromorphic systems results in systems with good
signal fidelity, high efficiency, high interconnectivity, highly parallel architectures, and high
information density. On the other hand, the integration of AI technologies into photonic
systems has proven to be promising for exacting knowledge from large datasets generated and
transported by photonic systems. The exploitation of machine-learning techniques has been
shown to have an impact in the areas of communications, imaging, and sensing.
Many challenges are still to be overcome and many diverse material platforms are being
investigated without a clear winner, the maturity of the developed hardware must improve (in
many demonstrators critical elements are still emulated in electronics), the scalability and the
robustness of the proposed PNN architectures should grow in order to meet the requirements
of typical machine learning problems.
Nevertheless the growing research efforts both from academia, prominent companies, and
fast- growing start-ups indicate that the coming years will witness a huge expansion of
this field, with strong potentials to address real-world applications.

AI for optical performance monitoring

A challenge in network control and management is to adapt to the time-varying link
performance parameters, such as optical signal to noise ratio (OSNR), nonlinearity factors,
chromatic dispersion (CD) and polarization mode dispersion (PMD). The progress of 5G
mobile networks, internet of things and cloud services has raised high demands and new
requirements for the capacity and reliability of optical networks.To fulfill the capacity demand,
low-margin optical network is attracting attentions. Therefore, planning tools with higher
accuracy are needed and accurate models for quality of transmission (QoT) and impairments
are the key elements to achieve this. Moreover, since the margin is low, maintaining the
reliability of the optical network is also essential and optical performance monitoring (OPM)
is desired. The estimation and acquisition of physical parameters of transmitted optical signals
allow network diagnosis in order to take actions (repairing damages, driving
compensators/equalizers or rerouting traffic around non-optimal links) against malfunctions.
With OPM, controllers can adapt the configuration of the physical layer and detect anomalies.
To improve the reliability of optical networks, controllers should be capable of obtaining the
real-time status of networks to prevent the serious degradation of systems.
To achieve this, advanced optical performance monitoring (OPM)
techniques are essential to enable needed functionalities to monitor the QoT and
impairments. If failures occur in optical networks, the monitoring mechanisms should be
capable of detecting, identifying and localizing them. However, considering the
heterogeneity of the modern optical network, it is difficult to build such accurate modeling
and monitoring tools using traditional analytical methods. Fortunately, data-driven artificial
intelligence(AI) provides a promising path.
Therefore, in summary, the modeling and monitoring techniques are
the key building blocks for the next generation EON and Artificial intelligence (AI)
technologies provide new opportunities to obtain a higher accuracy and/or a lower
complexity compared to analytical models.The basic architecture of the modeling and
monitoring techniques is shown:

JASMINE[17105051] Page 7
 Lec Assignment4 2020

For instance, Samples of received signals can be input to ML algorithms for monitoring the
chromatic dispersion (CD), polarization-mode dispersion (PMD) and optical signal-to-noise
ratio (OSNR) at the same time. Moreover, when obtaining information from the receiver
digital signal processing (DSP) modules, ML methods may be able to monitor the QoT or
impairments without any external devices such as the optical spectrum analyzer (OSA)
Therefore, for the next generation EON, applications of ML techniques for modeling
and monitoring can provide strong support to build a reliable and intelligent optical network
with lower design margins because:
 the monitoring information can guide the network self-reconfiguration and also
enables receivers to adapt some impairment compensation algorithms.
 Secondly, the real-time monitoring can continuously obtain the condition of the
physical layer.
ML methods are essential tools for network planning and management, but these methods
should be improved to be cost-effective and reliable for deployment.

It is experimentally observed that ML outperformed many traditional approaches for its

scalability, efficiency and robustness. In future, more research with ML will be carried out
for building an efficient, reliable and autonomous optical network. At the same time, there
are also some challenges for ML-based techniques for practical deployments.

Efficient adaptation scheme: For most of the works mentioned above, the ML-based
methods are trained offline with data from simulations or lab experiments before
deployment. Since the weights and parameters of the ML-based methods are fixed
after training, the calculation time will be short when using these methods in a
practical system. This firstly-trained-then-deployed scheme is efficient for adopting
ML-based methods for situations that require a fast response time. However, the data
from real scenes may be different from the simulation data. Therefore, a reasonable
adaptation scheme is also needed after deployment. In EON, online learning
approaches such as retraining are preferable to cope with time-evolving network
scenarios.

JASMINE[17105051] Page 8
 Lec Assignment4 2020
Reasonable design of ML structure: To reach a higher accuracy, ML algorithms with
more complex structures are introduced, such as DGCNN, reinforcement learning and
generative adversarial network (GAN). However, these ML methods with complex
structures may be hard to deploy in an optical system since they require large
memories. Therefore, cost-effective ML methods are desired for EON and the
structures of ML methods need to be adjusted to be tailored for the optical system.
Deployment of the ML engine: Many approaches for modeling and monitoring with
ML have been proposed recently. Where to deploy these ML engines is another
problem. Some ML engines can be embedded in receivers to build a low latency
system while some need to be deployed in the control plane to obtain information
from the whole optical networks Therefore, the strategies for the deployment of the
ML engine can be carefully designed to reach an optimum performance of the ML-
based method.

Role of AI for optical communications(either short or long reach):

Advance in optical transmission towards higher bit rate and denser spectral efficiency is
challenged by nonlinear effects. The tremendous growth of traffic requires continuing
advances in optical transmission systems to achieve higher capacity. Currently, the
information capacity of fiber optic systems is limited by nonlinear effects of the optical fiber.
Increasing the capacity while keeping the same reachability needs advanced techniques able
to mitigate linear and nonlinear transmission impairments .Extensive research effort has
attempted to address mitigation of nonlinearities on the transmission over optical fiber. Among
these nonlinearities, nonlinear phase noise (NLPN) is one of the prominent factors. Nonlinear
phase noise (NLPN) is the most common impairment that degrades the performance
of radio-over-fiber networks. The effect of NLPN in the constellation diagram
consists of a shape distortion of symbols that increases the symbol error rate due to
symbol overlapping when using a conventional demodulation grid.
So far this issue has been treated with electronic
methods relying on the deterministic information of the fixed fiber link, digital back
propagation and stochastic digital back propagation which may be computationally too heavy
for practical implementation. Compensation of linear impairments such as chromatic
dispersion (CD) and polarization-mode dispersion (PMD) are affordable using digital
equalization (e.g., fractionally spaced finite impulse response filter). However mitigation of
non-linear impairments induced by Kerr effect (i.e., self-phase modulation (SPM), cross-
phase modulation (XPM) and four wave mixing (FWM)) is still challenging.
Currently, artificial intelligence techniques are being incorporated to digital signal
processing to mitigate nonlinearities in a more efficient way, allowing more accurate symbol
detection. Neural networks are among the proposed solutions in literature to cope with the
complexity of those models. Taking advantage of the huge amount of data available in
optical transport networks, WDM systems represent a fertile field to apply neural networks.
The efficiency of NN to find nonlinear and complex relationships between input and output
data mainly depends on its conception (number of layers, number of neurons, etc.). In some
solutions, the conception of NN is derived from eNLE (eNLE-derived NN), while, in others,
NN is independent of the NLSE resolution (non-eNLE-derived NN).

JASMINE[17105051] Page 9
 Lec Assignment4 2020
The eNLE-derived NN consists either in describing an
eNLE, or determining physical parameters used to perform NLSE resolution. On the other
hand, non-eNLE-derived NN act as “black boxes”, regardless of optical line parameters.In
terms of complexity, NN computation is much lower compared to physical approach models
Furthermore, it is well-known that nonlinear equalizers such as DBP are especially efficient
to mitigate intra- channel nonlinear effects.In WDM configurations, their efficiencies are
limited, as inter-channel nonlinear effects involve physical mechanisms that need too heavy
computations for those equalizers.
Thus, Artificial Intelligence solutions focused on neural networks to mitigate fiber nonlinear
effects. NN solutions derived from existing nonlinear equalizers could help to cope with one
difficult step consisting in the determination of the number of layers, neurons and activation
function. NN offers a significant reduction in terms of computation complexity with
furthermore a slightly better system performance. Extending the proposed algorithms to
mitigate nonlinearity in WDM systems and dynamic environments could lead to find more
relevant scenarios to apply NN solutions.

CONCLUSION: Artificial Intelligence has been successfully applied to a wide

variety of problems in the context of telecommunication networks, mainly in wireless
networks, ranging from opportunistic spectrum access, to channel estimation and
signal detection in orthogonal-frequency division multiplexing (OFDM) to multiple-
input-multiple- output (MIMO) communications. Futher, AI presents several
opportunities for automating operations and introducing  in network planning and in dynamic
control and management of network resources, including issues like connection
establishment,  and self-optimization, through prediction and estimation by utilizing present
network state and historical data.

REFERENCES

 Shi, B.; Calabretta, N.; Stabile, R. Numerical Simulation of an InP Photonic

Integrated Cross-Connect for Deep Neural Networks on Chip. Appl. Sci. 2020, 10,
474.
 Shi, Bin & Calabretta, Nicola & Stabile, R.. (2020). Numerical Simulation of
an InP Photonic Integrated Cross-Connect for Deep Neural Networks on
Chip. Applied Sciences. 10. 474. 10.3390/app10020474.

JASMINE[17105051] Page 10
 Lec Assignment4 2020
 Z. Dong, F. N. Khan, Q. Sui, K. Zhong, C. Lu and A. P. T. Lau, "Optical
Performance Monitoring: A Review of Current and Future Technologies," in
Journal of Lightwave Technology, vol. 34, no. 2, pp. 525-543, 15 Jan.15,
2016, doi: 10.1109/JLT.2015.2480798.

 D. Wang et al.: Nonlinearity mitigation using a machine learning detector based on k-

nearest neighbors., IEEE Photonics Technology Letters, vol28.pp.2102-2105. 2016

 D. Zibar et al.: Machine learning techniques in optical communication, J.

Lightwave Technol, vol.34.pp.1442-1452, 2016

JASMINE[17105051] Page 11
JASMINE[17105051] Page 12
JASMINE[17105051] Page 13
JASMINE[17105051] Page 14