Newsletter

AI in Industrial Applications

Business benefits in real deployments, customer success stories

Interview with Jonathan Masci, Co-Founder and Director of Deep Learning

Hello Jonathan, in the many AI projects that you and your team have done over the years, which one was the most complex and rewarding?

Ciao Ralf, I must say that every project is special and rewarding in different ways, so this is a very hard choice for me. However, the User Assistance System that we developed for SCHOTT AG to improve their specialty glass production is definitely a milestone for industrial AI. From raw historical sensor readings we were able to build a 3rd generation digital twin (3DT) of the glass melting process. The 3DT was then embedded into a deep analytics engine where the expert operator, always in synergy with our AI, can gain further insights into the process and confirm reasons for historical process deviations for trust building. And, as the cherry on top, recommendations are provided to improve the desired business KPI by a targeted 10%. Why was this difficult? These processes are highly non-linear and exhibit non-stationary behavior: wear and tear and unobservable factors change the observed dynamics of the system throughout the course of its operation. In addition, due to the perfected production processes built over many years, historical data is repetitive and contains very little information about undesirable process dynamics. Interacting with the process to collect more data is costly and more importantly risky, which means our AI system had to learn from very few samples.

 

How would you compare your experiences in first projects that you did when you started back at IDSIA 10 years ago with the ones you do now with NNAISENSE?

I was always passionate about project management, and I surely did my best since early days at IDSIA. However, to be completely frank, when you are in academia, publishing and raising your scientific profile are key motivators. Simply put, you’re in love with your ideas and completely disconnected (disinterested) from the business aspect that is behind a specific industrial research area.  At NNAISENSE, we are driven by technology that works, is actionable, and provides our partners true benefits. The key is to understand where AI is essential, and that it is a cog in a very complex machine buy tiktok followers. Our role is to design technology that fits. In an academic setting, the research is its own end, whereas for us industrial AI is instead an enabler for better products and processes. Just having the technology adds little value, integrating it so that it can actually benefit your process is what really matters.

 

What type of challenges do you see in Industrial AI projects, and how do they differ from say an NLP-type project where one can use cloud-based off-she-shelf solutions?

The progress made in NLP in the recent years is unprecedented. It could actually generate the rest of this interview and you would not even notice, and perhaps it did. Behind this development are massive datasets and clever models like GPT-3 from OpenAI, but most importantly a shared semantics among all English written text. This makes it easy to standardize models and serve them on the cloud.  The same cannot be said of industrial processes. Each is different, and even those that are designed to be identical differ due to the impossibility of having identical physical realizations of the same process. This is the reason why no off-the-shelf solution exists, yet. More importantly, there are no massive datasets, and possibly there never will be in many applications due to the process being slow and/or dangerous to operate in certain regimes. It would simply take too much time and cost too much. At NNAISENSE, we are building tools to run industrial AI projects as conventional NLP projects are run now: Predictable results and smooth delivery with little risk involved. Our goal is to offer industrial AI solutions that work for the majority of cases, and that can leverage cross-domain knowledge to become better at solving new and more challenging problems over time.

 

What do you see as the most promising/interesting recent developments in deep learning, and how might they be applicable to industry?

The most promising development is certainly that of Graph Neural Networks (GNN) which are a type of deep learning model that can natively handle data that is not arranged on a grid, as is the case with conventional images, text, and videos. These methods are changing how computer vision is done, and enable faster and more successful drug discovery for example (see our recent paper on this), just to cite a few examples. In industrial AI, data is collected asynchronously from various sensors, with known spatial arrangement and dependencies, i.e. a sensor graph. A GNN can work on this data and leverage the structure of the domain as a strong regularizer to better generalize, for instance, in low data regime conditions.

The other area that is very relevant to achieving scalable industrial AI solutions is meta-learning, or learning to learn. Think of deploying a model in a factory and to being able to successfully transfer it to a (slightly) different factory almost seamlessly, without any further heavy lifting from your R&D team. This is also related to the emerging field of Offline RL, which we will cover in a future edition of our NNewsletter.

 

What are the differences between an AI project and a conventional engineering project?

A lot of what we do in AI involves engineering, of course. A good model gives you only a part, albeit essential, of the solution. However, the biggest difference with conventional engineering projects is that the trajectory to follow is often not clearly written a priori and must be discovered through an iterative process. Working with an experienced partner like NNAISENSE helps to mitigate this risk, to navigate close to optimal development. That being said, the industry also needs to go through a paradigm shift in how to approach AI projects, similar to what happened with agile software development. And last but not least, breaking with established patterns takes courage. If you want to get ahead of the competition, you need to be an early adopter of AI, and walk the rough path to success for better, more efficient and safer industrial processes.

Nnaisense News