.png){kind=icon}
.png){kind=icon}
Return
Efficient active learning using recursive estimation of error reduction
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2025.130131.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Efficient labeling is an open challenge in the big data era, where the rate of data collection far exceeds human capacity for data labeling required for training machine learning models. Active learning (AL) attempts to find the most valuable instances from a pool of unlabeled data by exploiting a limited annotation budget. In this paper, we focus on the estimated error reduction (EER) approach for AL. Given an initial labeled set, EER aims to identify the candidate unlabeled data points that, if used in training with other labeled data, minimize a weighted average of error estimates. As such, EER is a promising technique for AL because it directly examines the effect of labeling the candidate data points on the classifier's performance. However, the main shortcoming of EER is attributed to its excessive computational cost, as it requires retraining the classifier for each candidate data point and every possible label. In this regard, we propose an efficient EER algorithm using an accurate closed-form error estimator of Regularized Linear Discriminant Analysis (AL-RLDA) for the case of binary classification. At the same time, we introduce a novel objective function for EER, that relies less on the estimates of posterior probabilities. The efficiency of the algorithm is attributed to: (i) using a closed-form error estimator; and (ii) updating the estimator expression recursively for each candidate data point. Our empirical results using real data show that the proposed method can outperform baseline and state-of-the-art AL methods.
Keywords:
Active learning
Estimated error reduction
Regularized Linear Discriminant Analysis
Journal
IF:
6.5
Papers:
2.5W
Citations:
6.5W
Organization
No organization information available