notesum.ai

Published at May 12

Vector Quantization Prompting for Continual Learning

NeurIPS

Released Date: May 12, 2024

Authors: Li Jiao¹, Qiuxia Lai¹, YU LI², Qiang Xu³

Aff.: ¹Communication University of China; ²Harbin Institute of Technology, Shenzhen; ³The Chinese University of Hong Kong

Arxiv: https://openreview.net/pdf/fe56049dfd050804f643de97820660c0ab7ace62.pdf

Refer to caption

Method	Pub.	5-task		10-task		20-task
Method	Pub.	FAA ( $\uparrow$ )	CAA ( $\uparrow$ )	FAA ( $\uparrow$ )	CAA ( $\uparrow$ )	FAA ( $\uparrow$ )	CAA ( $\uparrow$ )
Joint-Train.		82.06		82.06		82.06
FT		18.74 $\pm$ 0.44	48.39 $\pm$ 0.58	10.12 $\pm$ 0.51	35.23 $\pm$ 0.92	4.75 $\pm$ 0.40	22.8 $\pm$ 0.37
FT++		60.42 $\pm$ 0.87	71.59 $\pm$ 0.50	48.93 $\pm$ 1.15	66.79 $\pm$ 0.92	35.98 $\pm$ 1.38	59.68 $\pm$ 0.95
L2P++ [62]	CVPR22	70.83 $\pm$ 0.58	78.34 $\pm$ 0.47	69.29 $\pm$ 0.73	78.30 $\pm$ 0.69	65.89 $\pm$ 1.30	77.15 $\pm$ 0.65
Deep L2P++ [62]	CVPR22	73.93 $\pm$ 0.37	80.14 $\pm$ 0.54	71.66 $\pm$ 0.64	79.63 $\pm$ 0.90	68.42 $\pm$ 1.20	78.68 $\pm$ 1.03
DualPrompt [61]	ECCV22	73.05 $\pm$ 0.50	79.47 $\pm$ 0.40	71.32 $\pm$ 0.62	78.94 $\pm$ 0.72	67.87 $\pm$ 1.39	77.42 $\pm$ 0.80
CODA-P [50]	CVPR23	76.51 $\pm$ 0.38	82.04 $\pm$ 0.54	75.45 $\pm$ 0.56	81.59 $\pm$ 0.82	72.37 $\pm$ 1.19	79.88 $\pm$ 1.06
HiDe-Prompt^* [59]	NeurIPS23	76.29 $\pm$ 0.10	78.77 $\pm$ 0.11	76.74 $\pm$ 0.18	78.76 $\pm$ 0.11	76.46 $\pm$ 0.06	78.76 $\pm$ 0.11
EvoPrompt [26]	AAAI24	77.16 $\pm$ 0.18	82.22 $\pm$ 0.54	76.83 $\pm$ 0.08	82.09 $\pm$ 0.68	74.41 $\pm$ 0.23	80.96 $\pm$ 1.42
VQ-Prompt	—	79.23 $\pm$ 0.29	82.96 $\pm$ 0.50	78.71 $\pm$ 0.22	83.24 $\pm$ 0.68	78.10 $\pm$ 0.22	82.70 $\pm$ 1.16