DeepART: Deep gradient-free local learning with adaptive resonance.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Deep neural networks excel at feature representation but struggle with lifelong learning.
• Adaptive Resonance Theory (ART) algorithms offer robust lifelong learning but can be computationally intensive.
• Integrating deep learning with ART can leverage the strengths of both paradigms.

Purpose of the Study:

• To introduce DeepART, a gradient-free, one-shot incremental learning technique for deep Hebbian neural networks.
• To adapt FuzzyART modules for deep network layers, enabling supervised learning via a FuzzyARTMAP head.
• To evaluate DeepART's effectiveness in lifelong learning scenarios with high-dimensional datasets.

Main Methods:

• Interpreting deep neural network layers as modified FuzzyART modules with specific input coding and weight update rules.
• Deriving local weight update rules for both fully-connected and convolutional layers.
• Utilizing a FuzzyARTMAP head for feature-category-label mapping to facilitate supervised learning.

Main Results:

• DeepART achieves a performance boost compared to existing ART-based methods.
• The technique effectively reduces category proliferation, enhancing scalability.
• Demonstrated improved processing of high-dimensional datasets in lifelong learning contexts.

Conclusions:

• DeepART successfully combines the feature learning of deep networks with the lifelong learning capabilities of ART.
• The gradient-free, one-shot incremental approach offers a scalable solution for complex learning tasks.
• This method presents a promising direction for advancing lifelong learning in artificial intelligence.