Dendrite Net: A White-Box Module for Classification, Regression, and System Identification.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Simulating biological dendrite computation is crucial for advancing artificial intelligence (AI).
• Existing machine learning (ML) algorithms like Support Vector Machines (SVM) and Multilayer Perceptrons (MLP) have limitations in biological system modeling.
• There is a need for ML algorithms that offer better interpretability and generalization capabilities.

Purpose of the Study:

• Introduce Dendrite Net (DD), a novel white-box ML algorithm inspired by biological dendrite computations.
• Evaluate DD's performance in system identification, regression, and classification tasks.
• Compare DD's generalization capability, accuracy, and speed against established ML architectures like Cell body Net (CBN).

Main Methods:

• Developed Dendrite Net (DD), a basic ML algorithm utilizing logical expressions (AND, OR, NOT) for classification.
• Tested DD on system identification tasks for black-box systems.
• Validated DD's performance on nine real-world regression applications and benchmark datasets (MNIST, FASHION-MNIST) for classification.
• Implemented and compared DD with Cell body Net (CBN) in MATLAB and PyTorch (Python).

Main Results:

• DD demonstrated excellent system identification performance for black-box systems.
• DD exhibited superior generalization capability in regression tasks compared to Cell body Net.
• For classification, DD achieved higher testing accuracy with greater training loss than Cell body Net.
• DD showed faster training (epoch) and forward propagation times than Cell body Net.

Conclusions:

• Dendrite Net (DD) is a white-box ML algorithm with controllable precision, offering enhanced generalization and lower computational complexity.
• DD's modular design allows for effective adjustment of logical expression capacity, preventing overfitting.
• DD presents significant potential for both generalized engineering applications and as a module within deep learning frameworks.