Interview with Anders L. Madsen, CEO of Hugin Expert — Why Bayes’ rule will never die

After this brief Easter break, we had another important announcement in the world of AI. Mark Zuckerberg, in the latest yearly F8, has revealed the long-term plan of Facebook, including AI and AR/VR as the main focus in their 10-year roadmap.

On December 19th, the CEO of Facebook had showcased his intelligent assistant, Jarvis, through a video on Facebook. It seemed only a sort of SCI-FI view of the world, but it is getting always closer and closer.

In 2027 more than 10% of the population might have flushed their smartphone and will be going around wearing some kind of AR “glasses”, if we want to try to give them a shape right now, empowered and enabled by AI.

Every day brings us closer to this new intelligent world.

This is the reason why we try every other week to give you the view of experts, entrepreneurs, and academics from the Nordic regions that live, breathe, and work in the amazing world of AI.

This week, we had a short talk with Anders L. Madsen, CEO of Hugin Expert, about his Aalborg-based technology company that has been working for the last two decades on Bayesian Networks, and on how you could use them to predict if the next Pope is going to be a black person.

As always I am craving to get your feedback or hear from you. Feel free to get in touch at

In addition, I want to announce our 4th AI Garage event in Copenhagen. This time Søren from Vivino and Rasmus from Tradeshift will be joining us! If you happen to be in Copenhagen, just stop by our office.

Could you give me an introduction about yourself and about what you guys do at Hugin Expert?

I am a computer scientist by training, graduated in 1996 at Aalborg University, one of the leading and pioneer university in the field of Bayesian Network and inference diagrams. I also received a PhD from Aalborg university in 1999 and joined Hugin Expert as a software developer right after that.

Today I am the CEO of the company and it has been so for 13 years. In addition to my PhD in decision support systems, I hold an MBA from Henley university in the UK.

As for Hugin, we are a small tech company and we work closely with Aalborg University. We have a very good relationship with them and try to take their technology and make it available to businesses, other universities and commercial entities.

How would you describe Bayesian networks to someone who has little or no knowledge about the topic?

A Bayesian network is a knowledge-integration tool deployed to manage uncertainty and support decision-making processes. It can also be considered as an intuitive graphical representation of dependencies relationship between different indicators and variables. These dependencies are quantified using probability, hence making possible to use Bayesian network to compute the probability of an event given evidences on the other variables studied, which could be both hard (data) or soft (expert knowledge) evidences. This is one of the best modeling tool for knowledge integration.

This is a huge advantage over other algorithms, because it allows the combination between expert knowledge and data into one single model.

A clear example is fraud detection. The steps that you need to go through in this case are the following.

1. Define the key indicators for the outcome you want to measure

The simple way to do fraud detection is based on the insurance company deciding on a set of features they usually measure in relation to a claim. Examples of these feature are the age of the person, how many claims the client has already had, how many policies the person has and so on. To these elements related to the client, you can combine indicators related to the claim in itself, as the damage claimed or the time and the location of the unfortunate event.

These indicators will become variables within the Bayesian network.

Then, the Bayesian network will also include a variable indicating whether or not the claim is fraudulent.

2. Structure the Bayesian Network and the various dependencies relationships

Given the various indicators and the final element to be measured, you can set up the whole structure of the Bayesian network and the definition of the dependencies relationships between the various indicators.

3. Quantify the joint probability of the dependencies

From there you can quantify the relations through probabilities, e.g how often is the person making the claim under 25 when you have a fraudulent claim. The calculation of this probability can either come from expert assessment or from the factual calculation of the percentage of fraudulent claims given that the person is under 25. In this case, you can combine both hard and soft data.

Simple example of a Bayesian Network — Here you can see the variables, the dependencies and the joint probability calculations.

Why have you decided to focus deeply on Bayesian Network and make it your key expertise?

We were born as a tool provider for Bayesian network. Indeed, Hugin was a spin off from a European project, focused on medical diagnosis problems related to muscles and nerves diseases. We could describe symptoms, disease and causes with a graphical model but there were yet no mean to calculate this probability. It was a gigantic joint probability distribution that Bayesian Network was perfectly fit to solve.

In parallel to this project, the founder of Hugin developed an algorithm able to solve this big graphical model involving joint probability distribution, which was applicable not only in a health care scenario, but more on a general basis. That’s why they decided to build Hugin, a company that should provide the tools to build Bayesian networks.

What are the main difference between Bayesian networks and random forest classifier? Why would you pick one over the another? And how DNN play a role in this game?

I am actually not an expert on random forest classifiers and DNN, but I think the Bayesian Network is a knowledge integration tool which has the huge advantage of being very easy to communicate to decision makers.

Indeed, you always have an intuitive graphical representation of the structure of the model. Even if it could be very difficult to build, it is always very easy to explain and communicate. This has been a great advantage in some domains, especially where models have to be approved and understood also by people which are not experts about the technology, as for example when dealing with modeling problems in healthcare or with financial institutions. The understandability of the model is a key differentiator.

Another big advantage is the possibility of combining expert knowledge and hard data directly in the models. This makes it possible to set up a full Bayesian Network without a single data case, something which is not possible with DNN. Indeed, you can drive the first part of exploration of the model only through expert knowledge.

A clear example is that a DNN cannot predict if the next pope is going to be a black person because we have a very few cases and we don’t have yet any black pope in the history of mankind. However, we could have an opinion that this could happen, especially thanks to experts in the field.

Building a neural network for this situation that had not happened yet is a challenge. On the other hand, if you have and expert opinion on that, you can set up your Bayesian network.

You have been working in the field for the last two decades, since AI and ML were not yet the sexiest topic in the world. How have you experienced the changes in the markets and the technology?

In the academic field, I am kind of active in two areas; the first one is hardcore Bayesian network, which aims to build models which are always better and faster. On this side, we have been experiencing a strong growth in terms of community and attention.

In the application-side of Bayesian Network, many more academics and PhD students use Bayesian Network in their research, independently from their specific area of expertise, to model problems for their particular field of studies.

AI-as-a-service. Is it reality or just a fantasy? Do you think that every single company can use and gain value from applying, without any knowledge and expertise, these AI algorithms to their data?

Right now, I currently see two different trends in terms of the use of our technology both from small and big companies.

First, we have companies using Bayesian Network as the driving technology of their business to solve problems for their final users. Examples that I can give you are efficient problem shooting or customer support.

On the other-hand, we have the chance to work with many companies that have a huge amount of data and want to use some tools to create knowledge from their data. In this case, they don’t need any proprietary algorithm, but only an out-of-the-box tool ready to be deployed.

So I am quite confident to state that AI-as-a-service can become a big thing.

Since 1998 you have also participated to the conference for the uncertainty in AI. What is the main uncertainty right now in the academic world focused on AI and ML?

Right now, I cannot see any obstacle to its development. AI is now being used by a large number of huge corporations and we will see an even wider adoption in the future.

I am not sure about what the barriers could be, but I am also inclined to say that I cant see any obstacle in the development of AI.

Recently you have also become Adjunct Professor of Computer Science at Aalborg University. How do you feel about this?

This is a great honorary title recognizing the scientific work that I have been doing with Aalborg university. This is another step to strengthen even more the bonds between the university and Hugin.

What is your favourite book?

It depends on who I am recommending to. I have recently read Black Swan, but the book that I would recommend is The Theory That Would Not Die. It is about Bayes’ rule and how it has been used in various settings. And, despite all the criticism, how it has survived 250 years. That would be my favorite one.

Where do you get your inspiration in terms of AI?

I think it is fundamental to be inside the research society, with close relationships with universities. Being part of the forefront of the European research is always exciting for me.

And what would you tell to entrepreneurs who want to use AI to disrupt new industries?

I think if I had some very good ideas I would have to shoot them myself. Set aside the joke, it is important to have an open mind and understand the real needs of the people, and only afterwards see how to solve them.

ML must not be developed for the sake of doing it, you always need to satisfy a need.