Directed Vectors for Generation of Independent Subspaces in the Bio-inpired Networks

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Machine learning

Background:

• Deep neural networks (DNNs) are widely applied but lack sufficient explanation for their complex structures and functions.
• Bio-inspired computing offers a promising avenue for understanding and elucidating DNN functionalities.
• Existing models often utilize symmetric architectures, potentially limiting explanatory power.

Purpose of the Study:

• To investigate the utility of bio-inspired asymmetric neural networks for explaining complex network functions.
• To compare the classification performance of asymmetric networks with traditional symmetric networks.
• To computationally demonstrate the generation of directional movement vectors and independent subspaces within asymmetric networks.

Main Methods:

• An asymmetric neural network was designed, drawing inspiration from biological retinal networks.
• The classification performance of the asymmetric network was evaluated against symmetric network counterparts.
• Directional vectors generated by adjacent neurons in response to movement stimuli were analyzed computationally.
• The creation of independent subspaces by these vectors and correlational activities was examined.

Main Results:

• Asymmetric networks, modeled after biological systems, were shown to be effective in explaining network functions.
• The asymmetric network demonstrated comparable or superior classification performance to symmetric networks.
• Computationally, directional movement vectors were generated in layered asymmetric networks, creating independent subspaces.
• Correlational activities among adjacent cells, represented as directed vectors, formed distinct independent subspaces compared to direct inputs.

Conclusions:

• Bio-inspired asymmetric neural networks provide a valuable framework for understanding the functional mechanisms of deep learning models.
• The generation of independent subspaces through directional vectors in asymmetric networks facilitates efficient sensory information processing.
• These findings suggest that asymmetric architectures can enhance feature extraction, classification, and learning in layered neural networks.