Learning Sudoku rules with conditional random fields
2017: EPFL - CVlab internship, under the supervision of Pierre Baqué and Pascal Fua
Combining backpropagation with Mean-Field inference models: the goal was to study these techniques in order to learn the rules of Sudoku as an ill-posed problem. Indeed given a grid it's possible to have multiple valid solutions. Pierre Baqué's (my internship mentor) Multi-Modal Mean-Field model should be able to cope with that kind of problems, combined with traditional machine learning techniques.
Learning the Sudoku rules as a set of constraints have been a success for the 4x4 version but I faced combinatorial difficulties in the 9x9 (traditional Sudoku) case, as inferring correct grids given a set of CRF constraints was proven to be a hard problem.