A degree of image identification at sub-human scales could be possible with more advanced clusters

While initial investigations in this domain primarily focused on the scaling of data volume, the present research takes a bold leap by not only scaling the data volume but also enhancing the image quality. This ambitious scaling experiment adopts a self-supervised learning approach that is both resource-efficient and feasible even without external financial support. Remarkably, our findings unveil the possibility of achieving human-level image identification performance at scales below human capabilities, achieved through the simultaneous scaling of data volume concentration and snapshot resolution. In order to do this, we conduct a scaling experiment using vision transformers and train them on a large data set with up to 200000 pictures, each of which has a resolution of 256 ppi.