Densely Connected Neural Networks for Nonlinear Regression.

Area of Science:

• Machine Learning
• Environmental Data Science

Background:

• Densely connected convolutional networks (DenseNet) excel in image processing.
• Convolutional DenseNet may lose crucial information in regression tasks due to independent input features.

Purpose of the Study:

• Propose a novel DenseNet regression model by replacing convolutional and pooling layers with fully connected layers.
• Maintain concatenation shortcuts to facilitate feature reuse.
• Investigate the impact of model depth and input dimensions on performance.

Main Methods:

• Developed a novel DenseNet regression model using fully connected layers and concatenation shortcuts.
• Conducted extensive numerical simulations to validate the model.
• Compared performance against support vector regression, decision tree regression, and residual regression.

Main Results:

• Identified an optimal model depth of 19 layers.
• Recommended limiting input dimensions to under 200 for optimal performance.
• The proposed DenseNet regression model outperformed baseline models at optimal depth.

Conclusions:

• The novel DenseNet regression model effectively reuses features and enhances regression task performance.
• The model shows high correlation when applied to predict relative humidity, indicating potential for environmental data science.
• The findings suggest this model can advance environmental data science applications.