Graduate Studies

I am a second year Masters student at McGill University, and my supervisor is Doina Precup. My main research interests currently revolve around machine learning and vision. My proposal to NSERC for scholarship purposes was Function Approximation in Large-Scale Tasks, which has been the focus of my studies so far. Here we are interested in developing methods for approximating value functions when they cannot exactly be solved. For many real-world applications, we cannot hope to have a small finite state space; this motivates the need for function approximation.

Current Research Interests

My research interests recently shifted towards a view of Artificial Intelligence geared towards sparse but strongly relevant experiences overlaid on a very rich but predictable world. Every day of our lives, we receive enormous amounts of information, or perceptions, from the outside world. On top of this we generate our own stream of consciousness. However, as adults we mostly ignore this information to concentrate on goals or more generally, uncommon events. My idea here is to develop a framework in which, once common occurences are accurately modelled, the system can focus on salient events. In such a case, temporal and spatial abstraction might be easier to construct as we can concentrate on things that truly need to be learned. I also strongly believe in the role of memory and the stream of consciousness mentioned above as a way of focusing learning. As such, I am interested in developping associative memories that can generate abstractions through relevance of concepts. Function approximation in theory provides the key to associative memories: we map a set of (perceptual) inputs to an associated item. However, hoping to obtain a good general function approximator when we can use a more restricted algorithm is not the simplest way one can develop associative memories.

Publications

• Joelle Pineau, Marc G. Bellemare, A. John Rush, Adrian Ghizaru and Susan A. Murphy, Constructing evidence-based treatment strategies using methods from computer science. Drug and Alcohol Dependence, 88, Supplement 2, 2007.
• Marc G. Bellemare, and Doina Precup, Context-driven predictions. Proceedings of the Twentieth International Joint Conference on Artificial Intelligence, 2007. [pdf]
  
• Marc G. Bellemare, Cascade correlation algorithms for on-line reinforcement learning. Proceedings of the North East Student Colloquium on Artificial Intelligence (NESCAI), 2006. [ps]