Comment by TheSaifurRahman

Has there been research on using this to make models smaller? If models converge on similar representations, we should be able to build more efficient architectures around those core features.