notesum.ai

Published at December 5

LossVal: Efficient Data Valuation for Neural Networks

cs.LG

Released Date: December 5, 2024

Authors: Tim Wibiral¹, Mohamed Karim Belaid², Maximilian Rabus², Ansgar Scherp¹

Aff.: ¹Ulm University; ²Dr. Ing. h.c. F. Porsche AG Stuttgart

Arxiv: http://arxiv.org/pdf/2412.04158v1

		Noisy Labels	Noisy Features	Mixed Noise	Overall Average
Classification	AME	0.074 $\pm$ .005	0.069 $\pm$ .005	0.069 $\pm$ .005	0.071 $\pm$ .003
	Beta Shapley	0.212 $\pm$ .003	0.191 $\pm$ .003	0.198 $\pm$ .003	0.201 $\pm$ .002
	DVRL	0.226 $\pm$ .005	0.187 $\pm$ .003	0.208 $\pm$ .004	0.207 $\pm$ .002
	Data Banzhaf	0.184 $\pm$ .004	0.162 $\pm$ .004	0.171 $\pm$ .004	0.172 $\pm$ .002
	Data-OOB	0.244 $\pm$ .005	0.186 $\pm$ .003	0.216 $\pm$ .003	0.215 $\pm$ .002
	Data Shapley	0.212 $\pm$ .003	0.191 $\pm$ .003	0.198 $\pm$ .003	0.200 $\pm$ .002
	Influence Subsample	0.184 $\pm$ .004	0.161 $\pm$ .004	0.170 $\pm$ .004	0.171 $\pm$ .002
	KNN-Shapley	0.355 $\pm$ .006	0.250 $\pm$ .005	0.298 $\pm$ .005	0.301 $\pm$ .003
	LAVA	0.099 $\pm$ .004	0.329 $\pm$ .012	0.220 $\pm$ .008	0.216 $\pm$ .005
	Leave-One-Out	0.173 $\pm$ .004	0.150 $\pm$ .004	0.166 $\pm$ .004	0.163 $\pm$ .003
	LossVal (epochs=5)	0.445 $\pm$ .007	0.213 $\pm$ .003	0.332 $\pm$ .005	0.330 $\pm$ .004
	LossVal (epochs=30)	0.544 $\pm$ .008	0.204 $\pm$ .004	0.371 $\pm$ .005	0.373 $\pm$ .005
Regression	AME	0.002 $\pm$ .000	0.002 $\pm$ .000	0.002 $\pm$ .000	0.002 $\pm$ .000
	Beta Shapley	0.002 $\pm$ .000	0.002 $\pm$ .000	0.002 $\pm$ .000	0.002 $\pm$ .000
	DVRL	0.247 $\pm$ .007	0.198 $\pm$ .004	0.218 $\pm$ .005	0.221 $\pm$ .003
	Data Banzhaf	0.002 $\pm$ .000	0.002 $\pm$ .000	0.002 $\pm$ .000	0.002 $\pm$ .000
	Data-OOB	0.002 $\pm$ .000	0.002 $\pm$ .000	0.002 $\pm$ .000	0.002 $\pm$ .000
	Data Shapley	0.002 $\pm$ .000	0.002 $\pm$ .000	0.002 $\pm$ .000	0.002 $\pm$ .000
	Influence Subsample	0.002 $\pm$ .000	0.002 $\pm$ .000	0.002 $\pm$ .000	0.002 $\pm$ .000
	KNN-Shapley	0.154 $\pm$ .008	0.111 $\pm$ .005	0.132 $\pm$ .006	0.132 $\pm$ .004
	LAVA	0.127 $\pm$ .004	0.415 $\pm$ .012	0.255 $\pm$ .008	0.265 $\pm$ .006
	Leave-One-Out	0.002 $\pm$ .000	0.002 $\pm$ .000	0.002 $\pm$ .000	0.002 $\pm$ .000
	LossVal (epochs=5)	0.380 $\pm$ .007	0.274 $\pm$ .005	0.330 $\pm$ .006	0.328 $\pm$ .004
	LossVal (epochs=30)	0.464 $\pm$ .008	0.256 $\pm$ .006	0.354 $\pm$ .006	0.358 $\pm$ .004