Dendrites endow artificial neural networks with accurate, robust and parameter-efficient learning

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Deep Learning (DL) algorithms, powered by Artificial Neural Networks (ANNs), excel at complex tasks but are parameter-heavy, energy-intensive, and prone to overfitting.
• Biological brains solve similar problems with remarkable efficiency, suggesting potential for bio-inspired AI.
• Current ANNs often require extensive training data and computational resources due to their architecture.

Purpose of the Study:

• To introduce a novel ANN architecture that mimics the structured connectivity and restricted sampling of biological dendrites.
• To investigate whether this dendritic ANN architecture can mitigate the limitations of traditional ANNs, such as overfitting and high parameter counts.
• To evaluate the performance of dendritic ANNs against traditional ANNs on image classification tasks.

Main Methods:

• Developed a new ANN architecture incorporating dendritic properties like structured connectivity and restricted sampling.
• Trained and evaluated dendritic ANNs on several benchmark image classification datasets.
• Compared the performance, parameter efficiency, and robustness to overfitting of dendritic ANNs against conventional ANNs.

Main Results:

• Dendritic ANNs demonstrated increased robustness against overfitting compared to traditional ANNs.
• The new architecture matched or surpassed the performance of traditional ANNs on image classification tasks.
• Dendritic ANNs utilized significantly fewer trainable parameters than conventional models, indicating greater parameter efficiency.

Conclusions:

• Incorporating biological dendritic properties into ANNs can lead to more precise, resilient, and parameter-efficient learning.
• The unique learning strategy of dendritic ANNs, where nodes respond to multiple classes, contributes to their advantages.
• This research highlights the potential of bio-inspired designs to enhance artificial intelligence capabilities and offers insights into ANNs' learning mechanisms.