The 29th Benelux Conference on Artificial Intelligence

As a fresh data scientist at Anchormen, I was pleasantly surprised to be sent to attend the Benelux Conference on Artificial Intelligence, which was being held In Groningen this year.

The conference was chaired by my thesis supervisor Dr. Marco Wiering and a lot of other familiar faces were present. The 29th edition of BNAIC was organized by the Institute of Artificial Intelligence and Cognitive Engineering (ALICE) from the University of Groningen, in collaboration with the Benelux Association for Artificial Intelligence (BNVKI) and the Dutch Research School for Information and Knowledge Systems (SIKS).

Because it was my first conference, I did not really know what to think. I must say, my expectations were exceeded! Especially the key note from a well know Facebook researcher, which you can read more about at the end of this blog.

Papers


The event was divided in three sections: Games, Knowledge & Reasoning, and Machine Learning. As a Data Scientist, the choice was obvious - Machine Learning all the way!

In the Machine Learning section four papers were presented. The first paper was about: “Simultaneous ensemble generation and hyperparameter optimization for regression”.

What stroke me the most was the use of Bayesian optimization instead of hand-tuning models. The second interesting part was that model generation can be done by combining multiple models in a single larger model. Among other topics, the author discussed the weighting, combining and dynamically growing these models.

The second paper was presented by a (computational) neuroscientist and one of my favorite teachers, Dr. Marieke van der Vugt. The paper is titled: “Tracking Perceptual and Memory Decisions by Decoding Brain Activity”. Dr. van der Vugt used intracranially recorded EEG to measure brain activity. In short, EEG measures brainwaves, which correspond to a certain mental process.

Intracranially means that the measurements are done within the skull (!)

She examined if memory and perception tasks use evidence accumulation to come to a decision. In other words, is it possible to decode decisions using brain activity. Parameterized faces were shown, and participants were asked if they were similar or not; she showed that it is indeed possible.

All the other papers can be found here.


Research and Business panel on A.I.


At two ‘o clock, our own Arjen van Wijngaarden represented Anchormen and sat down in a panel to discuss the next steps in Artificial Intelligence. Other business which were part of the panel, included NVIDIA, Target Holding, and Sim-CI. The main outcome of the panel discussion was to successfully incorporate data into businesses. Proof-of-concept needs to be created in order to show the value of the data as well as foster customer trust.

You can find out more about what Anchormen does with data and artificial intelligence here.


Keynote Lecture


The keynote lecture was from Dr. Laurens van der Maaten, a Research Scientist at Facebook A.I. Research in New York, working on Deep Learning, Machine Learning and Computer Vision.

His lecture was titled: “From Visual Recognition to Visual Understanding”. He explained how Deep Convolutional Neural Networks work and introduced a new/faster ensemble model.

It reminded me of the Scale Invariant Feature Transformation (SIFT) where pictures were augmented to different scales to increase recognition performance.

This is what Dr. van der Maarten did as well for his new multi-scale model. To speed up classification, he stacked models trained on data of different scale. New data is presented at the model classifying the smallest scale. When the output is higher than a threshold, models of higher complexity are not evaluated, thus achieving faster performance.

The most interesting part for me was recognizing bias in convolutional neural networks. In training models by asking questions, it could be that the questions were asked in such a way, bias would be introduced in the model. He used a famous horse named Clever Hans to explain his proposition. You can take a look at this interesting video here:

 
The well-known bias, probably known to most data scientists is that some classes may have a higher membership, and thus have a higher expectation value. He designed experiments which removed the bias by reformatting the experiment and got better results. He is an extraordinary intelligent individual, and I was very happy to be able to listen to him speak.


I had a great time and am looking forward to next year's BNAIC!  😊


Back to Data Science and AI