Stationary representations learned by a d-Simplex fixed classifier optimally approximate compatibility between sequentially fine-tuned and asynchronously replaced deep learning models.
