Unsupervised Visual Representation Learning via Dual-Level Progressive Similar Instance Selection.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Deeply learned representations require large-scale labeled datasets, which are often expensive or infeasible to obtain.
• Unsupervised methods are needed to overcome the limitations of data annotation in visual representation learning.

Purpose of the Study:

• To propose an unsupervised method for deep visual representation learning by leveraging instance discrimination and similarity.
• To introduce a dual-level progressive similar instance selection (DPSIS) method to enhance representation quality.

Main Methods:

• The proposed method utilizes instancewise classification, treating each instance as a unique class to learn meaningful representations.
• DPSIS adaptively selects two levels of similar instances (neighbors) for each anchor instance: 'absolutely similar' (hard labels) and 'relatively similar' (soft labels).
• The method progressively selects more neighbors without supervision as the convolutional neural network (CNN) model strengthens during training.

Main Results:

• Experiments on seven diverse benchmarks, including coarse-grained and fine-grained image classification datasets, demonstrate the effectiveness of DPSIS.
• The DPSIS method successfully improves deep visual representation learning by progressively identifying and utilizing similar instances.

Conclusions:

• Unsupervised learning using instance discrimination and similarity, particularly with the DPSIS method, is effective for deep visual representation learning.
• The DPSIS approach offers a viable solution for scenarios where large-scale labeled data is not available, significantly enhancing representation quality.