Pattern fusion in feature recognition neural networks for handwritten character recognition

Area of Science:

• Artificial Intelligence
• Machine Learning
• Neural Networks

Background:

• Neocognitron models face scalability issues due to large network sizes when handling numerous training patterns.
• Individual subnets for each training pattern in existing feature recognition networks limit efficiency.
• Combining features from similar training patterns can be problematic, potentially lowering recognition accuracy.

Purpose of the Study:

• To improve the efficiency and accuracy of feature recognition neural networks.
• To address the network size limitations of previous neocognitron-based approaches.
• To enhance the ability of neural networks to recognize patterns by effectively merging similar data.

Main Methods:

• Incorporation of fuzzy ARTMAP principles into a feature recognition neural network architecture.
• Development of a method to merge similar training patterns based on feature similarity.
• Implementation of shared subnets for fused training patterns.

Main Results:

• Significant reduction in overall network size.
• Demonstrated increase in the recognition rate.
• Improved handling of similar training patterns through feature fusion.

Conclusions:

• The proposed fuzzy ARTMAP-enhanced neural network offers a more scalable and accurate solution for feature recognition.
• Merging similar training patterns effectively addresses limitations of previous methods.
• This approach leads to more compact and higher-performing neural network models.