TY - GEN

T1 - Gaussian Based Active Learning Algorithm for Image Classification Problem

AU - Shcherbin, Andrey

AU - Yakhyaeva, Gulnara

N1 - Publisher Copyright: © 2021 IEEE.

PY - 2021/6/30

Y1 - 2021/6/30

N2 - One of the relevant problems in deep learning is data efficiency. In the active learning approach, we have a large set of unlabeled data, a small set of labeled data and a limited budget for labeling. The model for training on labeled data is defined. The task is to select the most relevant samples to increase model quality on the test data set. In this work, we overview some active learning algorithms and propose a novel algorithm for active learning based on Gaussian distribution. The main idea of the algorithm is to use reference samples from each class and compute the distribution parameters (and) for each embedding coordinate. We use Gaussian function as a measure of distance between unlabeled samples and class representation, so we can combine it with any confidence-based algorithm. We tested our approach on a part of ImageNet task (20 random classes from original ImageNet 2012 dataset). We used a Gaussian distribution-based measure combined with a confidence-based sample selection method. This combination achieved significantly better results.

AB - One of the relevant problems in deep learning is data efficiency. In the active learning approach, we have a large set of unlabeled data, a small set of labeled data and a limited budget for labeling. The model for training on labeled data is defined. The task is to select the most relevant samples to increase model quality on the test data set. In this work, we overview some active learning algorithms and propose a novel algorithm for active learning based on Gaussian distribution. The main idea of the algorithm is to use reference samples from each class and compute the distribution parameters (and) for each embedding coordinate. We use Gaussian function as a measure of distance between unlabeled samples and class representation, so we can combine it with any confidence-based algorithm. We tested our approach on a part of ImageNet task (20 random classes from original ImageNet 2012 dataset). We used a Gaussian distribution-based measure combined with a confidence-based sample selection method. This combination achieved significantly better results.

KW - active learning

KW - Gaussian filtering

KW - least confidence

KW - machine learning

KW - margin sampling

KW - maximum entropy

UR - http://www.scopus.com/inward/record.url?scp=85113484320&partnerID=8YFLogxK

U2 - 10.1109/EDM52169.2021.9507644

DO - 10.1109/EDM52169.2021.9507644

M3 - Conference contribution

AN - SCOPUS:85113484320

T3 - International Conference of Young Specialists on Micro/Nanotechnologies and Electron Devices, EDM

SP - 542

EP - 546

BT - 2021 IEEE 22nd International Conference of Young Professionals in Electron Devices and Materials, EDM 2021 - Proceedings

PB - IEEE Computer Society

T2 - 22nd IEEE International Conference of Young Professionals in Electron Devices and Materials, EDM 2021

Y2 - 30 June 2021 through 4 July 2021

ER -