Yocova recently interviewed Alexander Motzek, Principal Data Scientist at zeroG, a 100% subsidiary of Lufthansa Systems. Read this write-up about the evolution of artificial intelligence (AI), how zeroG is applying AI to benefit the wider aviation industry, and what Alexander believes the future will hold.
Alexander Motzek, zeroG
Alexander Motzek is the Principal Data Scientist at zeroG, which is a 100% subsidiary of Lufthansa Systems, based in Germany. He holds a PhD in Computer Science and Artificial Intelligence from the University of Lübeck, an MSc in Computational Engineering, and a BEng in Information and Electrical Engineering from Hamburg University of Applied Sciences.
Alexander drives AI within Lufthansa Group for 7 years and now works in the heart of artificial intelligence at zeroG. His main goals are to increase customer satisfaction by utilising the possibilities of AI, and to inform others about the benefits that can be derived from it.
In this article Alexander speaks to aviation journalist John Walton about his take on AI, how zeroG is using data and application programming interfaces (APIs), and the direction of travel for data and AI within the aviation industry.
Who is zeroG?
“zeroG is part of the Lufthansa Group and a 100% subsidiary of Lufthansa Systems. We are also working closely with Lufthansa’s ‘digital hangar’ and are driving topics mainly focused on data science, analytics, and artificial intelligence. Within that large ecosphere , our team works to rethink the complete aviation industry, through the lens of artificial intelligence. We try to bring all the new powers of AI, computer vision, deep learning, and neural networks to Lufthansa Group and the wider aviation industry.”
What is an API and what part do APIs play in modern aviation?
“Ultimately, APIs are like large documents defining protocols and they are protocols you can implement in any language. The great thing about them is that they are standardised protocols that can be implemented with just a few lines of code. If, for example, we want to interface with an application or provide an API based out of a Python application, we can simply import fast API and easily provide a function like analysing an image or text using a deep neural network by an API to any other application, say, even Excel.”
“APIs are a great enabler for us. Sometimes an API is simply a webpage where you can send information and get information back, which are so-called ‘rest APIs’. Then there are other APIs available in various forms where you can plug systems together and bring all their power together. This enables us to bring the huge power of artificial intelligence to link up systems, even legacy systems, into our AI solutions.”
How do you integrate legacy and modern languages?
“We can use standardised interfaces. For example, the HTTP(S) protocol enables us to communicate over the rest of the APIs and that has been standardised for decades, but what is happening behind is completely customisable. Take, for example, a simple webpage where you can upload an image, such as one taken at an airport gate. We can analyse what is happening at that moment, and basically identify things like exact aircraft position, if the gang way is docked, or if fuelling, baggage, and catering have arrived. And then by utilising multiple images or a video, we can derive complete situational awareness of each ongoing turnaround of aircrafts. This is our solution ‘DeepTurnaround’, which provides insights in real time for every aircraft turnaround and is used for process optimisations. This is all enabled through multiple APIs from our DeepTurnaround solution in combination with APIs from the Azure cloud.”
What are the biggest topics currently under discussion in the API sphere?
“This sphere has evolved over the past seven or so years and there have been some particularly interesting developments in recent months. Back in 2015, everything AI and API related was data analysis based on computer vision. Then around a year and a half ago these large generative AI models appeared that could draw pictures based on a short description you provide, which could create complete rooms and complete scenarios. That was probably the biggest buzz for around six years. Then, in December last year, this was all replaced by one big thing, which of course we all know of as ChatGPT. ChatGPT is an API – well, let’s be precise, first of all, it is a webpage. But behind that are very large language models, which you can access through an API, and you can ask it to summarise and analyse documents, texts and images or ask it questions.
“ChatGPT is not the only large language model available, and one of the things I love about it is the fact that we have been doing very similar things to that for years now. I created a large language model that could create text back in 2017, but that was 100,000 times smaller than ChatGPT. And then back in 2018, we trained a fully customisable model and made that available by an API that could analyse comments in respect of our catering services, and that was built on around 60,000 labelled data points. You don’t always need an API as large as ChatGPT. You can utilise small, very resource-friendly, and sustainable models for your desired application.”
What will this technology look like in future?
“It used to be that many systems were vendor locked and there was some secrecy, especially in terms of data itself, which was hidden in big silos. In the Lufthansa Group, we are driving a huge data literacy program which pushes us all to use the data we have available. I also see this trend in the wider aviation industry as well as in other industries such as automotive, where silos are being broken open. I think the cloud has contributed to this and data is now being made available and useful through APIs. It’s a fantastic time to be a data scientist because you finally have the tools available to create awesome new solutions.”
Has the discussion about who owns what data, and how, ended?
“I think data governance is an interesting topic as I often see many new rules popping up. The data owner is still there, and someone still owns the data, but it’s shifting a little bit in that data ownership is now more about data responsibility. There are new roles like data stewards coming into play, who curate the data in such a way that it becomes available in a much more useful and high-quality way. This is a really positive shift since very often the data owner comes from the respective business, so they know and understand the data and so can readily answer questions about it.
“We have some great insights with different data catalogs. It’s a nice piece of software you can use to search for anything, so it really breaks up the data silos and shows you exactly what’s inside the tables, where it comes from, who owns or stewards it and who to contact to access it.”
Should we just be following GDPR [EU General Data Protection Regulation] rules or another standard?
“In addition to GDPR there is also the EU AI Act coming up, which will state what is permitted to be done with data, the kind of artificial intelligence allowed to process this data and for what purpose. That is a huge step forward in terms of regulation in this arena. I think we should be careful not to over-regulate, but having regulations in place is very important since AI could be used for so many bad things. And crucially, it can be used badly by people who don’t know how to use it properly. It has become so easy to use APIs and feed it data and train it and then call the job done, but so much more care is needed than that. I come from a research background, and I know how incredibly hard it is to train an AI and then really prove that it is smart, that there is no glitch in the data, and that it has solved the problem it was supposed to. This responsibility needs to be in good hands, and it needs to follow strict scientific standards on how these models are evaluated. I have high hopes that the EU AI Act is going in that direction. Not just for the scientific community, but for the entire EU.”
Who are the big players when it comes to technology in aviation?
“At the moment, the big players are those who hold the data. If you don’t have any data, you can’t train anything and you can’t provide an API. You can build on some theories, but you need data to prove it. Having training data available is key to being a big player in the industry, and it helps us a lot. In the Lufthansa Group we have access to various data points, based on which we can train various models and build some awesome predictive maintenance solutions, for example, which can detect faults three months earlier than a maintenance engineer would notice them. We can build this independently because we have the complete expertise from academia, since that’s where we hire from. Obviously, we don’t have the capacity to compete with the ‘big guns’ like ChatGPT and Microsoft, neither financially nor from a sustainability perspective. But I would say that for around 95% of the use cases we have pursued over the past few years, we’ve not needed to use these large language models. There are many other more efficient, sustainable, and tailored solutions we can use and train ourselves.”
Which tech stack are you using for this and what expertise do you need to execute it?
“It’s divided really. You need good people to code, especially in Python deep learning models. However, this is becoming easier and easier, and the challenge now is how to keep up with all the academic developments of new models, but also with all the services evolving from the cloud, and ultimately how you can deploy everything you have developed and put it into production. For us at zeroG, teamed up with Lufthansa System’s data labs in Danzig and Budapest, we’re in a really good position now, from early proof of concept, to take the solutions right through to production readiness. Having this capacity is important to be able to utilise all the amazing potential out there. For this you need data engineers, data scientists and solution architects, and one of the most important things is to understand the business case. For this we employ business consultants to join up the different worlds and ensure things are not looked at in isolated ways.”
How will these technologies be used and what effect will that have on aviation?
“So, for us coders and data scientists, I hope we’ll be able to use tools like ChatGPT on a daily basis in a data conformed manner. Currently the way things are set up, data is stored somewhere, and I don’t want to use it in that fashion. If I’m assured data is not stored anywhere and is processed in a safe manner, then it becomes one of the most powerful tools we have seen on a level with programs like Word or Excel. Something that will help you with your daily work and help you code smaller snippets, and help you formulate text and help you draw ideas from a very big creative system. For example, in coding, there is a term called ‘rubber ducking’. Basically, if you’re stuck with a problem, you explain it to the rubber duck and by explaining it, you solve it. ChatGPT works this way, and it would be so helpful for our daily lives and a huge addition to our tech stack.
“In terms of aviation, I think for passengers there will be much more personalisation brought about by these tools. We’ll be able to improve the travel experience for all our passengers. Say, for example, damaged baggage – we’ll be able to send out messages with much more context and in a far more personal way, and this will be enabled by these recent developments in natural language, computer vision and reinforced learning. The initial results for this look really promising, so I’m excited about how we’ll move away from static messages towards dynamic messages that address you as an individual.”
What do you find the most fascinating part of all this new technology?
“I think the most fascinating part about AI is what you are capable of if you have the right data. The research community has been searching for the most efficient models and how to fine-tune and train algorithms to be smarter and come up with better decisions for a long time. Now, these large language learning models with deep learning have shown that if you throw enough data at it, it can solve the problem itself. We have tons of data available to use, but we need to find the right model to use and the right data to apply in a way that is efficient and also sustainable.”
Yocova on Air: Watch John Walton and Alexander Motzek discuss these topics in person.
To view the full podcast episode register for Free Yocova Membership, log in and head to Workspace > Learning.
Author: Yocova, with special thanks to John Walton
Published 01 August 2023
;
