notesum.ai

Published at November 27

ORIS: Online Active Learning Using Reinforcement Learning-based Inclusive Sampling for Robust Streaming Analytics System

cs.LG

Released Date: November 27, 2024

Authors: Rahul Pandey¹, Ziwei Zhu¹, Hemant Purohit¹

Aff.: ¹George Mason University

Arxiv: http://arxiv.org/abs/2411.18060v1

Refer to caption

Experiment	BERT Mini				mBERT
	Sigmoid		Exponential		Sigmoid		Exponential
	Machine	Human	Machine	Human	Machine	Human	Machine	Human
Random	40.4 $\pm$ 8.6	64.9 $\pm$ 11.2	47.3 $\pm$ 6.3	72.1 $\pm$ 5.4	47.0 $\pm$ 5.8	68.3 $\pm$ 6.5	46.8 $\pm$ 1.2	68.7 $\pm$ 2.9
Uncertainty	38.9 $\pm$ 2.4	62.7 $\pm$ 3.9	41.1 $\pm$ 2.7	65.3 $\pm$ 2.7	48.8 $\pm$ 6.4	68.4 $\pm$ 8.0	46.4 $\pm$ 9.2	67.7 $\pm$ 4.6
Diversity	51.0 $\pm$ 0	81.2 $\pm$ 0	36.1 $\pm$ 0	63.5 $\pm$ 0	50.4 $\pm$ 0	77.7 $\pm$ 0	50.4 $\pm$ 0	77.7 $\pm$ 0
ORIS ( $\delta=16$ )	54.4 $\pm$ 4.4	85.2 $\pm$ 4.6	52.7 $\pm$ 4.3	84.2 $\pm$ 6.1	54.1 $\pm$ 6.0	83.6 $\pm$ 8.2	55.6 $\pm$ 4.8	87.4 $\pm$ 2.9
ORIS ( $\delta=8$ )	56.2 $\pm$ 3.2	86.1 $\pm$ 4.6	56.2 $\pm$ 2.7	87.8 $\pm$ 4.0	60.6 $\pm$ 2.4	88.3 $\pm$ 3.6	58.5 $\pm$ 4.5	88.4 $\pm$ 3.1