Given the increasing importance of the neurosymbolic (NeSy) approach in artificial intelligence, there is
a growing interest in studying benchmarks specifically designed to emphasize the ability of AI systems
to combine low-level representation learning with high-level symbolic reasoning.

One such recent benchmark is Visual Sudoku Puzzle Classification, that combines visual perception with relational constraints.

In this work, we investigate the application of Logic Tensork Networks (LTNs) to the Visual
Sudoku Classification task and discuss various alternatives in terms of logical constraint formulation,
integration with the perceptual module and training procedure.