A not-too-technical explanation of Bayesian methods and machine learning can be found in >Dutch and English
Smart Research BV: Applications of neural networks and machine learning
Big4Data: our new commercial name. We do still Smart Research, but now we go Big4Data!
modeling Bayesian networks
Deterministic methods for
approximate inference in intractible graphical models (mean field, belief propagation, EP)
Multi-agent planning and control
Supermodeling! Applying machine learning to models with 100,000s of variables to improve modeling for climate science. A nice mix of chaos-theory, nonlinear dynamics, and machine learning. I participate in the STERCP project.
coordinated from the University of Bergen
SMART-RESEARCH BV is a spin-off company for commercial applications of neural nets,
Bayesian networks and other statistical modeling and reasoning techniques. I am executive vice-director.
Currently, we do quite well. Some recent projects that we are/were involved in are
Bayesian networks for mass disaster victim identification (in collaboration with NFI, partially supported by ICIS (see below)).
The system is called Bonaparte. Check its website with live demo
(for forensic professionals only, you need to register for the demo).
Bonaparte is now in use at Netherlands Forensic Institute (NFI), e.g. it has been used/is used
to identify victims from the Afriqiyah aircrash in Tripoli, Libya
to identify the victims from the MH17 disaster in Ukraine.
Some coverage of Bonaparte and MH17:
in the Marianne Vaatstra rape-murder case. The murderer was found in the biggest DNA dragnet search so far.
Recently a serial rapist in Utrecht has been caught.
He has been found in a familial search that was performed by using Bonaparte,
where he turned out to be the brother of somebody in the DNA database
NFI press release (Dutch)
. In the same way a serial killer has been caught (2017).
International users: Australia, Vietnam, Interpol.
For press releases, items in the news (e.g. NOS journaal item) and awards, see the Bonaparte website.
Other current or recent projects,
Power network analysis using MCMC
Prediction of individual employment potential
Diversity prediction. How to predict what costumers will put in their baskets to have a arrangement of items. Think of e.g. a holiday where you want a nice balanced mix of time spent in a village, a festival, a nice walk in nature, beach etc. I developed a method that deals with multi-item probability distributions.
Churn prediction (churn is when people change from telecom provider. Telecom companies want predict these risk-costumers, in order to make them a nice offer to stay. So be sure that you are a risk-costumer!!)
Wine advice: Neural networks and machine learning applied to learn the computer a taste for wine! Visit www.winewinewine.com, type your evening's dinner and find out what wine to drink!
Bovinose: detecting estrus in cow using an electronic nose. See FP7 site for a more detailed description.
Enose: cheap detection of tuberculosis using an electronic nose.
Fraude detection in on-line transactions.
Patterns in big data.
Optimization of machine settings.
Good old statistical analysis (consultancy based)
We are still proud of our
system for the prediction of newspaper sales, called JED, which is based on Neural Bayesian technology. It predicts
single-copy sales for individual outlets and also errorbars in this prediction. It can deal with missing values and it has
the ability that outlets learn from each other. Its main implementation has been at De TELEGRAAF, the largest Dutch newspaper. De Telegraaf saved several 100Keuros per year
by more efficient distribution of newspapers to the single-copy sales outlets.
Bayesian networks (BayesBuilder) have been applied a.o. for a system for bearing fault analysis Application (SKF), a decision support system for geotechnical engineering (SHELL), and for victim analysis (Bonaparte, see above).
NEWS: Big4Data is our new brandname under which we will develop all our new machine learning activities! We go Big4Data!
Pevious Scientific projects: SUMO
SUMO: climate modeling by Super Modeling. European project in collaboration with KNMI and other partners. 2010-2014. We were abels to demonstrate that quantitative improvements in climate modeling could be gained by connecting different climate models into one supermodel.
See EGU BLOG-POST. This project has a follow up:STERCP
Previous scientific projects: ICIS
I have been involved in the ICIS project.
See theirwebsite for more information on this huge research project. I did research on
multi-agent systems in continuous space and time. In our framework, we can employ
techniques from graphical models.
With Bart van den Broek (Ph-D student) and Bert Kappen. See our recent UAI papers.
Modeling of Bayesian networks, see BNAIC 2005 paper.
Application of hybrid MCMC methods in inference with continuous variables.
Application of Bayesian networks for forensic research (see above).
Some other previous projects
fMRI analysis using ICA (Donders)
I worked quite a long time on research and development of Promedas, a medical decision support system
based on a large Bayesian network. We developed this system in collaboration
with domain experts from University Medical Center Utrecht. The period of mainly research is over, we managed to create
systems of about 10000 nodes.
Now a company, Promedas BV, is founded whose aim is to commercialize the system. See
Promedas for more info. You can try the system on-line (!)
from this site.
SIENAStimulation Initiative for European Neural Applications -Here you can find a number of
successfull neural networks applications and some more info. Note: these are old pages,
and will not be not updated! In particular email adresses and links will be outdated. However, the examples on the site are still nice illustrations of neural network applications.
SIKS 2006: Introduction to Bayesian networks
MSc: Jonas Ahrendt (Pedigree reconstruction using MCMC), Christiaan Schoenaker (Bayesian attractor learning), Stefan van den Heuvel (Bayesian Optimization). BA: Joris Bukala (Importance sampling combined with MCMC).
for Neural Networks
Vereniging Artificiele Neurale Netwerken