Learning on tree architectures outperforms a convolutional feedforward network.

Area of Science:

• Neuroscience
• Computer Science
• Artificial Intelligence

Background:

• Current deep learning models with hundreds of layers are biologically implausible.
• Backpropagation's non-local weight changes are biologically unrealistic due to numerous pathways.
• A need exists for more biologically constrained deep learning architectures.

Purpose of the Study:

• To develop a biologically inspired 3-layer tree architecture for deep learning.
• To apply this architecture to the CIFAR-10 database for offline and online learning.
• To demonstrate an efficient dendritic deep learning approach.

Main Methods:

• A 3-layer tree architecture inspired by dendritic tree adaptations was developed.
• The architecture was applied to the CIFAR-10 database.
• A highly pruned tree backpropagation method was utilized, simplifying weight updates.

Main Results:

• The proposed architecture surpassed the success rates of the 5-layer convolutional LeNet.
• The dendritic tree architecture demonstrated effective offline and online learning.
• The pruned backpropagation approach proved efficient for deep learning.

Conclusions:

• The developed dendritic tree architecture offers a more biologically plausible deep learning model.
• This biologically inspired approach achieves superior performance compared to existing convolutional networks.
• Efficient dendritic deep learning is achievable through pruned backpropagation techniques.