Well, the approach currently doesn't apply to architectures being used in practice, and it's not clear, at least to me, if it can scale to large datasets.
Fair points. But given the potential impact this would have in practice, I'm still surprised there isn't more interest in exploring whether it's possible to scale the approach. Transforming neural network training into convex formulations would drastically simplify many current difficulties.
I'm baffled why Mert Pilanci's work in this area hasn't received more attention. His proofs of a zero duality gap for neural networks are impressive.