notesum.ai

Published at November 15

Forming Auxiliary High-confident Instance-level Loss to Promote Learning from Label Proportions

cs.AI

Released Date: November 15, 2024

Authors: Tianhao Ma¹, Han Chen¹, Juncheng Hu¹, Yungang Zhu¹, Ximing Li¹

Aff.: ¹College of Computer Science and Technology, Jilin University, China

Arxiv: http://arxiv.org/abs/2411.10364v1

Dataset	Model	Bag Size					Fully Supervised
Dataset	Model	16	32	64	128	256	Fully Supervised
CIFAR-10	LLPFC	84.10±0.19	71.70±0.78	52.71±0.36	20.78±0.70	18.79±0.21	96.05±0.33
	DLLP	91.59±1.52	88.61±0.90	79.76±1.45	64.95±0.01	44.87±1.13
	LLP-VAT	91.80±0.08	89.11±0.22	78.75±0.46	63.89±0.19	46.93±0.71
	ROT	94.86±0.68	94.34±0.65	93.97±0.96	92.23±0.81	63.10±0.84
	FLMm*	92.34	92.00	91.74	91.54	91.29
	L²p-ahil	94.96±0.13	95.00±0.11	94.58±0.21	93.64±0.20	92.88±0.53
CIFAR-100	LLPFC	-	-	-	-	-	79.89±0.14
	DLLP	71.28±1.56	69.92±2.86	53.58±1.60	25.86±2.15	8.82±0.94
	LLP-VAT	73.85±0.22	71.62±0.07	65.31±0.33	37.36±0.63	2.79±0.67
	ROT	72.74±0.08	69.31±0.22	17.48±0.86	11.02±0.79	2.86±1.11
	FLMm*	66.16	65.59	64.07	61.25	57.10
	L²p-ahil	78.65±0.28	77.30±0.50	76.52±0.23	72.21±0.37	23.56±2.13
SVHN	LLPFC	93.04±0.21	23.26±0.63	21.28±0.23	20.54±0.37	19.58±0.09	97.77±0.03
	DLLP	96.90±0.50	96.93±0.23	96.64±0.32	95.51±0.04	94.34±0.12
	LLP-VAT	96.88±0.03	96.68±0.01	96.38±0.10	95.29±0.17	92.18±0.29
	ROT	95.54±0.10	94.78±0.13	96.75±0.11	26.00±0.43	12.15±0.57
	FLMm*	-	-	-	-	-
	L²p-ahil	97.91±0.02	97.88±0.01	97.74±0.06	97.67±0.17	96.98±0.31
Fashion- MNIST	LLPFC	88.40±0.23	85.85±0.03	73.63±0.48	28.36±0.76	20.03±0.47	96.39±0.02
	DLLP	94.20±0.02	93.70±0.39	93.18±0.22	91.70±0.21	89.62±0.46
	LLP-VAT	94.69±0.20	94.17±0.16	93.25±0.18	92.30±0.13	89.51±0.51
	ROT	94.25±0.17	93.68±0.22	92.53±0.46	91.84±0.19	90.14±0.31
	FLMm*	-	-	-	-	-
	L²p-ahil	96.93±0.23	95.78±0.15	95.27±0.13	94.19±0.14	93.51±0.49
MiniImageNet	LLPFC	-	-	-	-	-	73.95±0.22
	DLLP	64.53±0.41	55.37±0.38	27.57±0.20	9.06±0.14	3.40±0.14
	LLP-VAT	64.17±0.34	54.36±0.29	30.96±0.24	9.69±0.17	4.90±0.09
	ROT	67.02±0.34	27.49±0.38	6.01±0.30	3.50±0.10	1.75±0.13
	FLMm*	-	-	-	-	-
	L²p-ahil	70.26±0.26	59.81±0.21	37.51±0.16	16.91±0.15	7.46±0.08