Figure 4.

The selection of transformations dictates the features learned during the training process, thus enabling the adaptation of a model for different tasks. A t-SNE projection of the ten-class clustering of the MNIST dataset^{( 29 )} was performed on two representations obtained from two self-supervised trainings of the same model using MoCo V2^{( 11 )}, with the sole distinction being the selection of transformations employed. One representation (a) retains information pertaining to the digit classes, achieved through padding, color inversion, rotation, and random cropping, while the other representation (b) preserves information regarding the handwriting font weight and style, achieved through vertical flips, rotation, random cropping.