R-ELMNet: Regularized extreme learning machine network

Area of Science:

• Machine Learning
• Computer Vision
• Deep Learning

Background:

• Principal Component Analysis Network (PCANet) is an effective unsupervised shallow network for feature learning.
• Existing Extreme Learning Machine Auto-Encoder (ELM-AE) variants, like ELMNet, have limitations in handling non-linearity or require complex preprocessing.
• There is a need for unsupervised feature learning methods that balance non-linearity and efficient projection.

Purpose of the Study:

• To analyze the intrinsic characteristics of ELM-AE variants.
• To propose a novel regularized ELM-AE that integrates non-linearity learning and approximate orthogonal projection.
• To evaluate the proposed method's effectiveness in unsupervised feature learning for image classification.

Main Methods:

• Developed a regularized ELM-AE by incorporating non-linearity and approximate orthogonal projection.
• Utilized a two-layer convolution structure inspired by PCANet.
• Conducted experiments on image classification tasks.

Main Results:

• The proposed regularized ELM-AE demonstrated effectiveness in unsupervised feature learning.
• Achieved competitive performance compared to supervised Convolutional Neural Networks (CNNs).
• Outperformed related shallow unsupervised learning networks.

Conclusions:

• The regularized ELM-AE offers a promising approach for unsupervised feature learning.
• The method effectively combines non-linearity and orthogonal projection for improved image classification.
• This work contributes to the advancement of shallow unsupervised learning techniques.