Non-Deep Active Learning for Deep Neural Networks.

Area of Science:

• Machine Learning
• Computer Vision

Background:

• Active learning enhances machine learning model accuracy by strategically selecting informative unlabeled data.
• Conventional active learning methods often rely on computationally intensive deep neural networks, leading to long training times.

Purpose of the Study:

• To develop a computationally efficient active learning method for image annotation.
• To improve the accuracy and speed of machine learning model training by optimizing data selection.

Main Methods:

• A non-deep neural network approach is proposed for unlabeled image selection.
• A task model is trained on labeled data to predict unlabeled images.
• An uncertainty indicator is generated based on model predictions to identify informative samples near the decision boundary.

Main Results:

• The proposed method achieves higher accuracy compared to conventional active learning techniques across multiple datasets.
• Execution time is reduced by up to 14 times (from 1.2 × 10^6 s to 8.3 × 10^4 s).
• Outperforms the current state-of-the-art (SoTA) method by 1% accuracy on the CIFAR-10 dataset.

Conclusions:

• The proposed method offers a significant improvement in both accuracy and efficiency for active learning.
• Selecting samples near the decision boundary is an effective strategy for informative data selection.
• This approach reduces the computational burden associated with active learning in machine learning.