64-channel correlator implementing a Kohonen-like neural network for handwritten-digit recognition

Area of Science:

• Artificial Intelligence
• Optical Computing
• Pattern Recognition

Background:

• Kohonen maps (self-organizing maps) are unsupervised learning algorithms for dimensionality reduction and visualization.
• Optical implementations offer potential for high-speed parallel processing in neural networks.
• Handwritten digit recognition is a critical task in postal automation and data entry.

Purpose of the Study:

• To present an optical implementation of an improved Kohonen map neural network.
• To apply the system to the recognition of handwritten digits from a postal code database.
• To demonstrate the benefits of supervised learning in simplifying map layer labeling and improving performance.

Main Methods:

• An improved Kohonen map neural network architecture was developed.
• Supervision was introduced during the learning stage to enhance performance and simplify labeling.
• A frequency-multiplexed raster computer-generated hologram was used for N(4) interconnections.
• The optical setup was characterized as a 64-channel correlator.
• Computer simulations were performed to evaluate detection and classification strategies.

Main Results:

• The optical system successfully recognized handwritten digits from a postal code database.
• Experimental results obtained using binary phase computer-generated holograms showed excellent agreement with computer simulations.
• The introduction of supervision led to simplified map layer labeling.

Conclusions:

• The presented optical implementation of a supervised Kohonen map neural network is effective for handwritten digit recognition.
• Computer-generated holography provides a viable method for realizing complex interconnections in optical neural networks.
• The study validates the potential of optical computing for efficient pattern recognition tasks.