An improved Artificial Protozoa Optimizer for CNN architecture optimization.

Area of Science:

• Artificial Intelligence
• Computational Biology

Background:

• Neural Architecture Search (NAS) is crucial for optimizing Convolutional Neural Networks (CNNs).
• Existing optimization algorithms can suffer from premature convergence, limiting exploration of potential CNN architectures.

Purpose of the Study:

• To propose MAPOCNN, a novel NAS method utilizing an enhanced Artificial Protozoa Optimizer (APO).
• To improve CNN architecture optimization by mitigating premature convergence and enhancing solution exploration.

Main Methods:

• Developed Modified Artificial Protozoa Optimizer (MAPO) incorporating protozoa phototaxis behavior.
• Applied MAPOCNN to optimize CNN architectures on benchmark datasets like Rectangle and Mnist-random.

Main Results:

• MAPOCNN demonstrated faster convergence compared to existing NAS methods.
• The proposed method achieved competitive performance, outperforming others in speed and accuracy.
• MAPOCNN effectively explores a wider range of CNN architectures.

Conclusions:

• MAPOCNN offers an efficient approach to discovering high-performing CNN architectures.
• Biologically inspired optimization techniques show promise for advancing deep learning architecture design.