Evolution of Deep Convolutional Neural Networks Using Cartesian Genetic Programming.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Deep learning models like Convolutional Neural Networks (CNNs) excel in computer vision.
• Designing complex CNN architectures demands significant expertise and extensive trial-and-error.
• Automating CNN architecture design is crucial for advancing AI capabilities.

Purpose of the Study:

• To develop an automated method for constructing high-performing CNN architectures.
• To reduce the reliance on expert knowledge and manual design in CNN development.
• To achieve competitive performance on computer vision tasks with optimized computational efficiency.

Main Methods:

• Utilized Cartesian genetic programming (CGP) to encode CNN architectures.
• Employed functional modules like convolutional blocks and tensor concatenation as CGP node functions.
• Optimized CNN structure and connectivity via evolutionary algorithms to maximize accuracy.
• Implemented techniques for accelerated architecture search, including rich initialization and early network training termination.

Main Results:

• Achieved competitive performance against state-of-the-art models on CIFAR-10 and CIFAR-100 datasets.
• Demonstrated the ability to find effective CNN architectures with reasonable computational cost.
• Showcased a significant reduction in computational time and resources compared to other automated design methods.

Conclusions:

• The proposed automated CNN architecture design method is effective and efficient.
• Cartesian genetic programming provides a viable framework for optimizing deep learning models.
• This approach offers a practical solution for developing advanced computer vision systems with reduced resource demands.