Dendritic normalisation improves learning in sparsely connected artificial neural networks.

Area of Science:

• Computational neuroscience
• Machine learning

Background:

• Artificial neural networks (ANNs) are inspired by biological neurons.
• Sparsely-connected ANNs offer computational efficiency and biological architectural resemblance.
• Tuning connectivity in sparse ANNs is an active research area.

Purpose of the Study:

• To introduce a novel normalization technique for ANNs based on neuronal dendritic biophysics.
• To evaluate the effectiveness of this dendritic normalization in various sparse network architectures.

Main Methods:

• Developed a normalization method dividing artificial neuron afferent contact weights by their number.
• Applied this dendritic normalization to sparsely-connected feedforward, recurrent, and self-organized networks.
• Compared performance against other widely-used normalization techniques in sparse networks.

Main Results:

• Dendritic normalization significantly improved learning performance across tested sparse network architectures.
• The proposed method outperformed existing normalization techniques in sparse network settings.
• Achieved practical advancements in machine learning and provided insights into dendritic computation.

Conclusions:

• Dendritic normalization is a highly effective technique for enhancing sparse artificial neural network performance.
• This method offers a practical improvement for machine learning applications.
• The study provides valuable insights into the computational role of neuronal dendritic structures.