While this is certainly a issue for large datasets, I wonder if this could be leveraged in some way to make better use of small data. For instance, consider you only have 5 classes with 20 images on each. Here, the n³ could actually be a way to learn something useful from so little data.