Self-organizing maps on "what-where" codes towards fully unsupervised classification.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Unsupervised learning is gaining traction due to the high cost and biological unnaturalness of large labeled datasets.
• Current unsupervised methods often rely on supervised classifiers, necessitating prior knowledge of class numbers and labels.
• Self-organizing maps (SOMs) show promise as unsupervised classifiers but typically require high-quality embeddings from deep learning techniques.

Purpose of the Study:

• To develop a completely unsupervised, end-to-end classification system using a Hebbian approach.
• To integrate a previously proposed What-Where encoder with a SOM, eliminating the need for labels or prior class knowledge.
• To demonstrate a system capable of online training and adaptation to emerging classes.

Main Methods:

• Utilizing a What-Where encoder in tandem with a Self-Organizing Map (SOM).
• Implementing a Hebbian learning rule for end-to-end unsupervised classification.
• Conducting experimental analysis on the MNIST and Fashion-MNIST datasets.

Main Results:

• The proposed system achieves accuracies comparable to state-of-the-art methods on the MNIST dataset.
• The system demonstrates robust performance on the more challenging Fashion-MNIST dataset.
• Validation of the system's ability to perform classification without requiring labels or pre-defined class information.

Conclusions:

• The integrated What-Where encoder and SOM form an effective, end-to-end unsupervised Hebbian classification system.
• This approach overcomes limitations of traditional methods by not requiring labels or prior knowledge of class numbers.
• The system's online adaptability suggests potential for real-world applications with evolving data distributions.