Competitive repetition suppression (CoRe) clustering: a biologically inspired learning model with application to...

Area of Science:

• Computational Neuroscience
• Machine Learning
• Artificial Neural Networks

Background:

• Determining compact neural coding is crucial for understanding biological memory and artificial neural networks.
• Optimal network structure remains an open problem in unsupervised learning.
• Repetition suppression (RS) is a cortical mechanism generating compact visual stimulus representations.

Purpose of the Study:

• Introduce a novel learning algorithm, competitive repetition-suppression (CoRe) learning, inspired by cortical RS.
• Derive a clustering algorithm (CoRe clustering) that automatically estimates cluster numbers without prior data information.
• Analyze CoRe learning dynamics and its relationship with existing clustering models.

Main Methods:

• Developed the CoRe learning algorithm based on the repetition suppression mechanism.
• Derived the CoRe clustering algorithm from the general CoRe learning model.
• Provided an error function to describe CoRe learning dynamics for analysis.

Main Results:

• CoRe learning generates compact neural representations of visual stimuli.
• CoRe clustering automatically estimates the unknown cluster number from data.
• CoRe clustering demonstrates robustness to noise and outliers.
• CoRe learning extends rival penalized competitive learning (RPCL) by enhancing penalization using robust statistics loss functions.

Conclusions:

• CoRe learning offers a biologically plausible approach to compact neural coding.
• CoRe clustering provides an unsupervised method for automatic cluster number estimation.
• The CoRe model offers theoretical advantages in robustness and extends state-of-the-art methods.