notesum.ai

Published at November 18

Continual Task Learning through Adaptive Policy Self-Composition

cs.AI

Released Date: November 18, 2024

Authors: Shengchao Hu¹, Yuhang Zhou¹, Ziqing Fan¹, Jifeng Hu², Li Shen³, Ya Zhang¹, Dacheng Tao⁴

Aff.: ¹Shanghai Jiao Tong University, Shanghai AI Laboratory; ²Jilin University; ³Sun Yat-sen University; ⁴Nanyang Technological University

Arxiv: http://arxiv.org/abs/2411.11364v1

Refer to caption

Benchmarks		OCW10			OCW20
Metrics		P ( $\uparrow$ )	F ( $\downarrow$ )	FWT ( $\uparrow$ )	P ( $\uparrow$ )	F ( $\downarrow$ )	FWT ( $\uparrow$ )
MT	MTL	$0.75\pm 0.04$	-	-	$0.80\pm 0.02$	-	-
Reg	L2	$0.29\pm 0.06$	$-0.01\pm 0.00$	$0.03\pm 0.05$	$0.20\pm 0.01$	$0.00\pm 0.00$	$0.01\pm 0.01$
	EWC	$0.16\pm 0.02$	$0.67\pm 0.02$	$0.02\pm 0.03$	$0.12\pm 0.02$	$0.69\pm 0.02$	$0.05\pm 0.01$
	MAS	$0.29\pm 0.04$	$0.47\pm 0.07$	$0.01\pm 0.01$	$0.25\pm 0.02$	$0.44\pm 0.02$	$0.01\pm 0.02$
	LwF	$0.21\pm 0.04$	$0.62\pm 0.03$	$0.06\pm 0.04$	$0.10\pm 0.01$	$0.70\pm 0.05$	$0.06\pm 0.03$
	RWalk	$0.26\pm 0.03$	$0.01\pm 0.01$	$-0.02\pm 0.01$	$0.17\pm 0.01$	$0.05\pm 0.02$	$0.00\pm 0.01$
	VCL	$0.14\pm 0.03$	$0.68\pm 0.04$	$0.04\pm 0.03$	$0.06\pm 0.00$	$0.74\pm 0.04$	$0.04\pm 0.01$
	Finetuning	$0.11\pm 0.03$	$0.73\pm 0.04$	$0.03\pm 0.01$	$0.08\pm 0.00$	$0.77\pm 0.03$	$0.03\pm 0.02$
Struc	LoRA	$0.54\pm 0.03$	$0.00\pm 0.00$	$0.01\pm 0.00$	$0.54\pm 0.01$	$0.00\pm 0.00$	$0.02\pm 0.02$
	PackNet	$0.64\pm 0.06$	$-0.01\pm 0.01$	$0.03\pm 0.01$	$0.57\pm 0.04$	$0.00\pm 0.00$	$0.05\pm 0.02$
Reh	PM	$0.26\pm 0.01$	$0.56\pm 0.05$	$0.04\pm 0.03$	$0.26\pm 0.08$	$0.57\pm 0.09$	$0.11\pm 0.02$
	A-GEM	$0.12\pm 0.04$	$0.70\pm 0.06$	$0.02\pm 0.01$	$0.09\pm 0.01$	$0.73\pm 0.04$	$0.06\pm 0.02$
Ours	Grow	$0.60\pm 0.06$	$-0.01\pm 0.01$	$-0.01\pm 0.01$	$0.61\pm 0.02$	$0.01\pm 0.02$	$-0.01\pm 0.03$
	Prune	$0.69\pm 0.01$	$-0.01\pm 0.03$	$0.00\pm 0.00$	$0.73\pm 0.04$	$0.00\pm 0.01$	$0.03\pm 0.02$