Self-supervised learning on smaller datasets is not feasible to translate into larger datasets like ImageNet, and the learning problems in multimodal cases create different challenges for scaling up.