Co-Training for Visual Object Recognition Based on Self-Supervised Models Using a Cross-Entropy Regularization.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Deep learning excels at visual object recognition but requires extensive labeled data.
• Acquiring large labeled datasets is costly and time-consuming.
• Semi-supervised learning, like co-training, offers a solution by utilizing limited labeled data.

Purpose of the Study:

• To propose a novel co-training model for visual object recognition.
• To reduce the dependency on large labeled datasets in deep learning.
• To enhance the efficiency and accuracy of visual object recognition systems.

Main Methods:

• Developed Differential Self-Supervised Co-Training (DSSCo-Training) model.
• Integrated self-supervised neural networks as intermediate inputs.
• Diversified data views using cross-entropy regularization of outputs.
• Employed deep neural networks for visual object recognition.

Main Results:

• The DSSCo-Training model demonstrated competitive performance against state-of-the-art methods.
• Achieved an average relative accuracy improvement of 5% across multiple datasets.
• Validated on well-known image datasets: MNIST, CIFAR-100, and SVHN.
• Showcased superior performance despite its simpler architecture.

Conclusions:

• DSSCo-Training effectively merges co-training and self-supervised learning principles.
• The model offers a simpler yet effective approach to visual object recognition.
• Presents a promising direction for efficient deep learning with limited labeled data.