Skip to main content
. 2024 Nov 14;4:e12. doi: 10.1017/S2633903X2400014X

Figure 3.

Figure 3.

Transformation choices influence superclass performance. We analyze mean superclass accuracy in Cifar100 using BYOL, SimCLR, and VICReg SSRL methods, varying crop size (a) or hue intensity (b). Our observations show consistent patterns across models, highlighting distinct effects of transformation parameters on different superclasses. Each superclass has unique optimal parameters, underlining the ability of transformation selection to modulate superclass performance.