ChaosNet: A chaos based artificial neural network architecture for classification.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• The brain's chaotic neural firing inspires novel computational models.
• Existing artificial neural networks often require extensive training data.
• Chaotic maps like Generalized Luröth Series (GLS) have demonstrated utility in data compression and cryptography.

Purpose of the Study:

• To propose ChaosNet, a novel artificial neural network architecture for classification tasks.
• To develop a learning algorithm for ChaosNet that leverages the properties of chaotic neurons.
• To evaluate ChaosNet's performance and robustness, especially with limited training data.

Main Methods:

• ChaosNet architecture utilizes layers of 1D chaotic maps (Generalized Luröth Series - GLS).
• A new learning algorithm is designed, exploiting the topological transitivity of GLS neurons.
• Performance is assessed on public datasets for classification tasks, including evaluations with minimal training samples.

Main Results:

• ChaosNet achieves high classification accuracy (73.89%-98.33%) even with very few training samples per class (less than 0.05% of data).
• The network demonstrates robustness against additive parameter noise.
• A two-layer ChaosNet implementation shows enhanced classification accuracy.

Conclusions:

• ChaosNet offers a promising approach for classification tasks, particularly when training data is scarce.
• The proposed learning algorithm effectively utilizes the properties of chaotic GLS neurons.
• Future research may explore additional learning algorithms for ChaosNet.