Formation of a direction map by projection learning using Kohonen's self-organization map

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Neurons in the primary visual cortex exhibit selectivity for edge orientation and motion direction.
• Some visual cortex areas display complex organization, including subdivisions for opposite motion directions within iso-orientation domains.
• The input signal space for direction maps in visual cortex is often non-convex.

Purpose of the Study:

• To address the limitations of standard Self-Organizing Maps (SOMs) with non-convex input spaces.
• To introduce a modified SOM algorithm incorporating projection learning.
• To validate the efficacy of the modified SOM for analyzing complex neural data.

Main Methods:

• Modification of Kohonen's Self-Organization Map (SOM) algorithm.
• Introduction of a projection learning method to constrain reference vectors to the input signal space.
• Application of the modified SOM to ferret and cat visual cortex direction maps.

Main Results:

• The projection learning method successfully constrains SOM reference vectors to the input signal space, even when non-convex.
• The modified SOM accurately reproduced a direction-orientation joint map from ferret and cat visual cortex data.
• This approach enables the analysis of non-convex input signal spaces, overcoming a key limitation of standard SOMs.

Conclusions:

• The proposed projection learning method enhances the applicability of SOMs to non-convex input signal spaces.
• This modified SOM is a valuable tool for analyzing complex neural representations, such as direction-orientation maps.
• The study demonstrates the successful modeling of visual cortex organization using advanced machine learning techniques.