Ph.D. student in machine learning (under the supervision of Prof. P. Geurts)
Systems and Modeling Research Unit
Department of Electrical Engineering and Computer Science
University of Liège
Contactjm.b...@ulg.ac.be +32 (0) 4 366 29 72 I 127 Montefiore (B28), Sart-Tilman
Montefiore Research Unit
University of Liège
Montefiore Institute (B28)
I am interested in tree-based representation learning for topologically-structured data.
In high dimensional classification problems, such as image classification (ex. computer vision) and time series classification (ex. speech recognition), the main difficulty is finding interesting features (i.e. a Representation Space) for the task at hand. Doing so automatically, a scheme known as Reprensentation Learning allows for more generic algorithms, faster developments and broadens the field of potential applications.
Representation Learning is usually associated with neural networks and other deep learning achitectures portraying hidden layers. In my thesis, I am considering how classification forests can be used for that purpose.
Vast as the field of Representation Learning is, I have decided to restrict the subject to Topologically-structured data. That is, data composed of homogenous elements related by a neighborhood relationship.
Indeed, many learning algorithms operating on those data do not take (enough) into account those relationships. Take a look at the picture: it is a "4" whose pixels are gradually shuffled. Once the pixels are shuffled enough we are unable to recognize the digit. However, this do not affect many learning algorithms (provided the shuffling is the same for all images). Put in other words, those algorithms are capable of learning a model of something which is definitely not a "4". The performances (namely the classification rate) are not impacted because the algorithms cannot take into account topolical relationships. Conversely, if they were able to capture that information ; if they understood what a "4" really is, they would make less mistakes at recognizing them.
By focusing on learning topologically-rich representations, the hope is to be able to raise the performances of tree-based techniques so as to benefit from their natural advantages (speed, scalability, interpretability, ease-of-use, etc.)
|2014 - present||
Ph.D. student in Machine Learning, Univertisy of Liège
Advisor: Prof. P. Geurts
|2011 - 2014||
Master degree in Computer Science, University of Liège
Master thesis: Generic image classification: random and convolutional approaches
under the supervision of Prof. P. Geurts and Ph.D. R. Marée
Awarded of the Melchior Salier Prize "Best Master's Thesis in Computer Science" (A.I.M)
|2009 - 2011||
Master degree in Geomatics and Geometrology, University of Liège
Master thesis: Raster data server prototype: Rasdaman
under the supervision of Prof. J.P. Donnay
|2006 - 2009||Undergraduate degree in Geographic Science, University of Liège|
|2005 - 2006||Cambridge English : Advanced (CAE) in Malta|