Completely Automated CNN Architecture Design Based on Blocks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional Neural Network (CNN) performance is heavily dependent on architecture design.
• Designing effective CNNs requires significant expertise in both the field and the specific problem domain.
• This expertise is not universally available, posing a barrier to entry for many users.

Purpose of the Study:

• To develop an automated method for evolving CNN architectures.
• To eliminate the need for user expertise in CNNs, problem domains, or genetic algorithms (GAs).
• To create a system that requires no preprocessing or postprocessing steps.

Main Methods:

• Utilized a genetic algorithm (GA) for automatic CNN architecture evolution.
• Incorporated ResNet and DenseNet blocks within the GA framework.
• Evaluated the algorithm on CIFAR10 and CIFAR100 benchmark datasets.

Main Results:

• The proposed GA-based algorithm automatically designs CNN architectures.
• Achieved superior classification performance compared to hand-crafted and fully automatic state-of-the-art methods.
• Demonstrated competitive accuracy against semi-automatic approaches.
• Consumed significantly fewer computational resources than most competing methods.

Conclusions:

• The automated GA-based approach effectively designs high-performing CNN architectures.
• This method democratizes CNN design by removing the need for specialized knowledge.
• Offers an efficient and resource-conscious alternative for developing advanced CNN models.