Neural networks with multiple general neuron models: a hybrid computational intelligence approach using Genetic...

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Classical neural networks rely on pre-specified neuron models and fixed architectures.
• Existing models often involve complex weight and bias tuning, limiting adaptability.
• The need for more flexible and automatically evolving neural network architectures is evident.

Purpose of the Study:

• To present a new type of neural network, termed NN-GP, that autonomously learns neuron functions.
• To demonstrate a flexible network architecture where any interconnection is possible.
• To explore a model that eschews traditional weights and biases, simplifying the learning process.

Main Methods:

• Utilized Genetic Programming (GP) to evolve general functions for individual neurons.
• Implemented an Evolutionary Computation process to search the function space for optimal neuron models.
• Evaluated network performance using fitness measures based on classification or regression errors.

Main Results:

• Developed NN-GP networks capable of learning unique neuron models.
• Achieved flexible network layouts without reliance on connection weights or biases.
• Demonstrated promising performance on real-world classification problems through dimensionality reduction.

Conclusions:

• NN-GP offers a novel approach to neural network design by evolving neuron functions.
• The method facilitates the creation of low-dimensional representations from high-dimensional data.
• This evolutionary approach holds potential for advancing classification and regression tasks.