Nonlinear Regression via Deep Negative Correlation Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Nonlinear regression is crucial for computer vision tasks like crowd counting and age estimation.
• Current deep learning solutions involve robust loss functions or deep network ensembles, each with limitations.

Purpose of the Study:

• To propose an efficient "divide and conquer" method for deep nonlinear regression.
• To address the limitations of existing deep learning approaches in regression problems.

Main Methods:

• Generalizing negative correlation learning for deep regression.
• Systematically controlling the bias-variance-covariance trade-off without extra parameters.
• Creating an ensemble of accurate and diversified deep regression models.

Main Results:

• The proposed method yields deep regression ensembles where base models are both accurate and diversified.
• Sub-problems exhibit reduced Rademacher Complexity, facilitating easier optimization.
• Demonstrated superiority and versatility across crowd counting, personality analysis, age estimation, and image super-resolution tasks.

Conclusions:

• The "divide and conquer" approach offers an efficient and effective solution for deep nonlinear regression.
• The method successfully balances bias-variance-covariance for improved performance.
• Validated through extensive experiments on diverse and challenging computer vision benchmarks.