notesum.ai

Published at December 3

FCL-ViT: Task-Aware Attention Tuning for Continual Learning

cs.AI

Released Date: December 3, 2024

Authors: Anestis Kaimakamidis¹, Ioannis Pitas¹

Aff.: ¹Aristotle University of Thessaloniki

Arxiv: http://arxiv.org/pdf/2412.02509v1

Refer to caption

	10 tasks			20 tasks			50 tasks
Methods	#TP	Avg	Last	#TP	Avg	Last	#TP	Avg	Last
iCaRL [59]	11.22	65.27 $\pm$ 1.19	50.74	11.22	61.00 $\pm$ 2.03	43.75	11.22	56.08 $\pm$ 3.08	36.62
WA [81]	11.22	69.44 $\pm$ 1.29	53.78	11.22	67.33 $\pm$ 1.13	47.31	11.22	62.08 $\pm$ 1.08	42.14
DER [76]	11.22	74.64 $\pm$ 0.65	61.72	11.22	75.09 $\pm$ 0.84	62.48	11.22	75.01 $\pm$ 0.90	61.23
DyTox	10.73	74.02 $\pm$ 1.04	60.67 $\pm$ 0.14	10.74	72.27 $\pm$ 0.88	56.32 $\pm$ 1.06	10.77	70.00 $\pm$ 1.45	52.34 $\pm$ 0.26
FCL-Vit	14.23	77.63 $\pm$ 0.42	65.02 $\pm$ 0.05	14.23	76.60 $\pm$ 0.31	67.61 $\pm$ 0.12	14.23	78.59 $\pm$ 0.56	67.72 $\pm$ 0.15