This work studies a learning framework where neural architecture is not fixed in advance but emerges during training. Learning is formulated as a trajectory on a curved information manifold, with an entropy field guiding the evolution toward stable configurations.