Computational modeling ...

My principal research interest consists in building computational cognitive models of human learning from empirical data. I apply machine learning tools to account for specific human skill acquisition, and more specifically for unsupervised concept learning. The main problem I am interested in is to investigate to which extent it is possible to predict the representations that are more likely to emerge from a given set of stimuli.

The goal is to study general principles underlying cognitives processes and implement them through a unique computer program. It is then applied to different tasks in order to test its range of validity. The validation is done by comparing results obtained by the model and by human participants when exposed to the same set of stimuli.

... to search for fundamental cognitive principles ...

I am currently focusing my research on the simplicity principle as a general cognitive mechanism. Theoretical results based on Algorithmic Information Theory suggest that simplicity can be used as a selection criteria to create models with a high predictivity power. It could be seen as an application of Occam's razor to the field of cognition. Using Information Theory, the intuitive notion of simplicity could be formalized by the Minimum Description Length principle (MDL), thus allowing a compression of statistical regularities. Cognition can be seen as a compressor that is able to extract information from incoming stimuli in order to create a "compressed" representation of the external world. When shaping perception, some patterns are more likely to be extracted. In my thesis, I investigate the simplest one called "chunk".

The aim is to investigate what kind of representations are the most likely to emerge using this simplicity principle as a selection criteria. To test this hypothesis, that the representations created by cognitive agents are those allowing the shortest encoding of input stimuli, I designed two programs and compared their predictions against human data.

... and compare them against human data.

The first model called MDLChunker is a raw implementation of the simplicity principle applied to model the way simple representations (chunks) are acquired by humans. MDLchunker is based on MDL and on the notion of chunking broadly used in cognitive psychology. This parameter-free model is able to create new chunks by simplifying (in the MDL sense) the input dataset. This model successfully applied to three different tasks:

• Learning word segmentation from a continuous speech stream (linguistic field)
• Modeling short term memory capacity in a visual memorizing task. Capacity could be expressed in terms of a fixed information quantity (in bits) instead of a fixed number of elements (chunks).
• Learning new concepts (chunks) in a simple drag and drop task. The model is able to predict the chunks created along with the time course of chunk creation.

This model is able to create new representations of its environment by extracting statistically relevant regularities, which in turn influence the perception of future data (perception shaping).

The second model called MDLChunker-cog is based on the same principles but is designed to account for cognitive limitations. It uses an architecture similar to a neural network, thus allowing parallel information processing using local information only.