Evolving Connections in Group of Neurons for Robust Learning

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Traditional neural networks utilize hierarchical multilayer architectures.
• These networks lack connections between nodes within the same layer.
• Deep learning models have shown significant success in machine learning.

Purpose of the Study:

• To propose a novel group architecture for neural network learning.
• To develop a probabilistic model for optimizing neural network connections.
• To enhance the robustness of neural networks, particularly against corrupted data.

Main Methods:

• Neurons are irregularly assigned within groups, allowing flexible connections.
• A probabilistic model optimizes connections based on input-output node relevance.
• Particle swarm optimization is employed to evolve network architecture directly, bypassing weights and biases initially.

Main Results:

• The proposed architecture demonstrates superior performance compared to existing models on noise-corrupted images.
• The model achieves this performance even when trained solely on clean images.
• The architecture is found to be robust against data corruptions.

Conclusions:

• The novel group architecture offers significant improvements in machine learning tasks involving noisy data.
• Directly modeling architecture without initial weights and biases reduces computational complexity.
• This approach provides a more robust and efficient alternative for neural network design.