Table 3.

The results of the adjusted mutual information score ^{( 47 )} obtained for two sets of transformations, with different SSRL approaches and backbones, through the mean of five training runs for each, compared to each other and to the AMI score achieved on the representations of pre-trained models (Resnet 101 and VGG16) trained with supervision on ImageNet, and applied on the dataset subsets containing Nocodazole, Cytochalasin B and Taxol. The selection of the pretrained models width is studied in Supplementary Materials. Both sets of transformations comprise random rotations, affine transformations, color jitter, and flips, with the first set including an additional random cropping, and resulting in a mediocre AMI score, and the second set applying random rotations and resulting in a significantly higher score.

Transformations	SSRL approach	Backbone	Nocodazole	Cytochalasin B	Taxol
First set :	MoCo v2	VGG13	0.19	0.27	0.16
		ResNet18	0.17	0.25	0.15
	Byol	VGG13	0.21	0.28	0.19
		ResNet18	0.2	0.25	0.17
	VICReg	VGG13	0.19	0.26	0.2
		ResNet18	0.16	0.25	0.21
Second set :	MoCo v2	VGG13	0.37	0.45	0.38
		ResNet18	0.33	0.42	0.31
	Byol	VGG13	0.38	0.48	0.41
		ResNet18	0.35	0.44	0.34
	VICReg	VGG13	0.38	0.44	0.36
		ResNet18	0.34	0.43	0.3
Pretrained models on ImageNet		VGG16	0.34	0.55	0.36
		ResNet101	0.39	0.57	0.43