notesum.ai

Published at November 4

ViTally Consistent: Scaling Biological Representation Learning for Cell Microscopy

cs.LG

cs.AI

cs.CV

68T07

I.2; I.4

Released Date: November 4, 2024

Authors: Kian Kenyon-Dean¹, Zitong Jerry Wang¹, John Urbanik¹, Konstantin Donhauser², Jason Hartford³, Saber Saberian¹, Nil Sahin¹, Ihab Bendidi², Safiye Celik¹, Marta Fay¹, Juan Sebastian Rodriguez Vera, Imran S Haque¹, Oren Kraus¹

Aff.: ¹Recursion; ²Valence Labs; ³University of Manchester

Arxiv: http://arxiv.org/abs/2411.02572v1

Model backbone	$b$	CORUM	hu.MAP	React	StringDB	KS	CVM
Baseline ViTs
ViT-S/16, Untrained	$12$	.45	.34	.205	.36	.30	4.3
ViT-S/14, Dino-V2	$12$	.48	.345	.20	.38	.34	5.6
trimmed	$5$	.51	.36	.21	.40	.35	6.0
ViT-L/16, ImageNet WSL	$24$	.52	.35	.21	.39	.34	5.5
ViT-L/14, Dino-V2	$24$	.49	.34	.21	.38	.34	5.3
trimmed	$12$	.55	.37	.22	.41	.36	5.9
ViT-L/16, ImageNet MAE	$24$	.53	.355	.215	.40	.34	5.1
trimmed	$11$	.53	.36	.22	.40	.35	5.8
ViT-G/14, Dino-V2	$40$	.44	.31	.20	.35	.29	3.8
trimmed	$16$	.53	.35	.22	.40	.33	5.2
MAEs for microscopy
CA-MAE-S/16 , RxRx3	$12$	.55	.37	.23	.43	.47	10.4
MAE-L/8 , RPI-93M	$24$	.61	.43	.25	.47	.52	12.3
trimmed	$15$	.60	.43	.255	.475	.57	15.2
MAE-L/8 , PP-16M	$24$	.60	.43	.255	.48	.59	16.2
trimmed	$20$	.60	.435	.26	.48	.59	16.2
MAE-G/8 , PP-16M	$48$	.62	.44	.26	.49	.60	16.4
trimmed	$38$	.615	.44	.26	.49	.63	18.2