The approach is solid, but the choice of learning rules, circuit architecture, and sparsity assumptions is not always sufficiently justified, and the model would benefit from a clearer mechanistic ...