Nicolas Dupuis and AI application in Telco industry

Published by Leila Rebbouh on

It’s been a little over a year that I discuss with Nicolas Dupuis on LinkedIn. We thought we should meet but it never happened. With the launch of LegIA Squad, it seemed obvious to me that we could not miss this great name in Artificial Intelligence, a guy from Liège moreover !


Nicolas Dupuis, DMTS Senior Machine Learning Scientist,
AI Strategy and Innovation Lead,
Nokia Software

Winner of the two last Global AI Conferences organized by Nokia Bell Labs, Distinguished Member of Technical Staff and main author of 30+ patents in the field of Broadband communication & AI, Nicolas Dupuis is Lead for AI Innovation & Machine Learning developments within Nokia Software. His strong domain expertise and his long-term contact with the market led him to create pioneering product features globally deployed at major service providers. Nicolas studied both in Belgium and in the Netherlands, holds a M.Sc. in Electrical Engineering and is a Member of both the European AI Alliance and the IEEE ML for Communications ETI.

Leila : Hello Nicolas. Can you explain us your professional background?

Nicolas: Right after graduated from a M.Sc. in Electrical Engineering (“Ingénieur civil Électricien, orientation Électronique”) in 2008 from ULiège and having enjoyed my last year abroad applying some machine learning within the medical Signal Processing Systems group at the Eindhoven University of Technology (TU/e), I naturally got attracted by a position within Thales Alenia Space to deal with signals and to do some advanced data mining.

From 2010, I joined Alcatel-Lucent, world-leader in Wireline network technologies, as a subject matter expert within a software product line. I strongly dig into the physical layer of communication systems (e.g. signals, waves, code correction, propagation, …) and their implication toward the end-user’s quality of experience. As co-creator of diagnosis software globally deployed at major service providers and operating on country-wise networks, I extensively travelled on the 5 continents to support team involved at any stages of the product life-cycle.

From 2016, after Alcatel-Lucent got acquired by Nokia to become the world-leading company in Wireline, Wireless and Core network infrastructure (with a portfolio covering all that we do not see from the Internet, ranging from optical Core routers to xDSL street cabinets passing by 4G/5G antenna and cloud-native management/optimization software), my main focus has been on the transformation of our feature design process into a machine learning driven approach as well as on the creation of concrete and business-relevant AI innovation.  These developments have been awarded as winner of the two last global AI conference organized by Nokia Bell Labs, reaching today some network-wide deployment of our deep-learning based systems within several service providers live networks.

Leila : You are now working at Nokia. Can you explain your role to us?

Nicolas : Even if I am part of a large-scale company, world-class products may originate from few people within confined teams. Within Nokia, some of these teams are located in Belgium. In that context, my current role is very heterogenous. Of course, it implies to think and propose innovative technical features that leverage AI as its best to solve today and tomorrow telco service providers’ problematic. But, as a Lead, this also implies to mentor direct and indirect colleagues, to drive cross-functional program/projects, to support presales and advise product managers, to visit customers, to deal with C-level executives, to advocate externally about our AI innovation, for instance by giving talks at major conferences and, last but not least to stay accurate and to finally turn these ideas into practice, to have my hands myself on TensorFlow, Keras, Spark and Python code on regular basis. At the end, we are not far from the Start-up world, however taking advantages of a bigger structure and a recognized brand (e.g. marketing, sales channel, resources,…).

Leila: You work mainly in ‘remote’. What are the advantages and disadvantages of this way of working? How do you overcome the disadvantages?

Nicolas: Of course there are pros- and cons-. Working remotely allows an extreme flexibility, sometimes far from the 9-17h schedule we can be used to. It also allows to reach some level of focus hardly feasible in open space offices. By contrast, it also implies a lot of discipline, autonomy and good coordination/communication/trust between the different team members. However, having at least once a week some face-to-face meeting and some video conferences in between remain essential.

Leila: You’re in charge of a team with more than 30 patents. It’s really impressive, especially since you’re relatively young. How do you explain this success? Is it due to the corporate culture or have you recruited very specific talents?

Nicolas : Even if this seems a lot, it represents about 3 to 4 patents per year since 2008, which is finally not so huge, even for a small team. About the secret sauce, this can be explained by a combination of factors. First, when you operate or become an expert in a niche sector, you can rapidly see that there is not especially a huge prior art. In that context, a good idea – usually a specific algorithmic feature – can be more easily patented. Second, it does not need to be extremely complex to be patented. Something that is novel and that works effectively may lead to a patent. Also, Nokia has a well-established process for patenting concrete developments (Nokia is making >1 billion USD annually from the revenue of its active patents). This also facilitates the filing. About the people – but this is linked to my first point – there is usually no talent available for a niche market. The key is to hire smart people with strong scientific bases, to train them with some specific background and to mentor them properly so that the idea we can get implemented more effectively, leading ultimately to patent.

Leila : During our meeting, you made a very clear distinction between Data Science and Artificial Intelligence. Can you tell us more ?

Nicolas : Indeed. My main interests and my focus is now on building systems – whatever it is via classical machine learning or advanced deep learning – that “learns” from data and take advantage of this. It is finally like being in a teacher role. As a teacher, there is a significant part of the job that consists in selecting the right topics, the right books, on the right matters, focusing on the right chapters, etc. This is similar with machines. Creating the training set from which the system will learn from is one of the major aspects and, to do so, there is nothing better than creating yourselves your data. This could also be very advantageous for the time-to-market of a product, as there is no need to wait for massive deployment before getting data. A good example are the robots from Boston Dynamics, for instance, that have been trained on synthetic data, as there is no field data available as they are not so many real robots in the streets. But they are working pretty well. In that sense, they are living examples that a machine can learn relevant information being trained from synthetic data.By contrast, learning directly from raw field data rapidly introduce some complexity that we do not want at this stage. Moreover, this could limit the control, the completeness and the variability that would be required to suitably train our system, leading at the end-to lower performances. The labelling might also be unclear or simply not present. So, Data Science suits best to extract insights from a data base but not specifically to derive a model from it. It is in that sense that I am making the distinction with a Machine Learning-based functionality, addressing at the end another purpose.

Leila: So you’re applying artificial intelligence to the Telecom sector. Do you think it is better to be an artificial intelligence specialist first or a telecoms specialist first to work in such a specialized sector?

Nicolas: That depends. Important are the fundamentals. Or where it is faster to ramp up in knowledge. To be able to understand where AI can bring value for some specific aspects of the telecommunications technology, it is crucial to understand deeply these technologies first. And to understand them, their parameters, their dependencies, the big pictures but also the deeper details… it is usually a long-term process, whatever the business (Telecommunications, Aerospace, Genomics,…). In that context, I have observed that such specific domain expertise was key. Or saying differently, that it would have taken more time for a data scientist to acquire all that domain knowledge than for a domain expert to get familiar with some machine learning concepts. But for less specific problematics (i.e. where data science knowledge gets more time to acquire than the subject to apply on), I would probably say and advise the opposite, as going deeper into some deep learning aspects would also require some experience.

Leila:  What advice would you give to a recently graduated young data scientist if he wanted to follow in your footsteps?

Nicolas : Become an expert in something and focus on what matters. For that, choose a domain and try to go as deep as you can in it. Whatever it is telecommunication, Data Science, Imaging, Oncology,… that does not matter. Try to be coached by thought leaders that focus on the right aspects and do not get disturbed by the “bling bling”. When Yann LeCun managed to put in practice the Convolutional Neural Networks he had been trusting in for years when he was at (Nokia) Bell Labs in the 1990’s, he was far from being popular. These days you cannot talk about Deep Learning without referring to him, moreover knowing he got the Turing Award and is regularly on covers of broad audience magazines.

Leila : Do you know the #AI4Good initiative? Are there certain principles that you apply in your sector of activity?

Nicolas : The Unicef targets? Yes, of course. As a Finnish company, Nokia cares and is engaged to reduce its CO2 emissions for instance. Moreover, it is at the agenda to make the telecommunication sector less energy greedy. Via our latest technologies and software – either by better configuration or more efficient processes – we directly contribute to these reductions. At the end, all these “AI-driven” innovations acting on these aspects contribute to consume less, hence applying these AI4Good considerations to our sector.

More generally, being part of one of the most ethical company worldwide, I am specifically engaged to develop AI where it helps human and, of course, not where it bypasses humans. I am regularly in contact with field practitioners who deal with or consume the results of our AI. Even if they had fears in the early days, these professionals are now our best supporters as they clearly see the improvements in their daily job, whatever their skills or education level, and our AI is also used to train them !

Leila : Last question. Beyond your sector of activity, how do you see the evolution of artificial intelligence?

Nicolas : From a technical point of view, we are still (very) far from the human intelligence. It is not because a machine can learn by processing millions of examples that – in fine – the machine is “intelligent”. That is particularly visible when there is a need for extrapolation. With a couple of pictures from cats and dogs, a child can easily infer the presence of ears on an elephant. Today this is far to be as simple for machines. I see therefore a need for complementing the today’s approaches with methods that require less training data and with models that would be able to extrapolate/generalize better. There is emergence of “semi-supervised” techniques as well as neuronal models 3.0 (or even 4.0), getting closer from biological ones. Adding more contextual information would also help.

By contrast, I see that the level of accuracy and reliability reached by AI in some dedicated tasks is already overtaking the one of humans. In that context, and without entering into the explainability subject that is also popular these days, better to address how we can live with AI and take advantages of it than to go against it.