Structure and Base Analysis of Receptive Field Neural Networks in a Character Recognition Task

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Shallow convolutional neural networks (CNNs) with fixed kernels are explored, particularly concerning training with limited data.
• Research on Receptive Field Neural Networks (RFNN) is extended to analyze network behavior with architectural modifications.

Purpose of the Study:

• To investigate the impact of different bases and architectural changes on RFNN performance.
• To establish a reproducible methodology for RFNN training and evaluation.
• To evaluate the significance of architectural changes using Bayesian comparison.

Main Methods:

• Simplified a baseline RFNN to a single-layer CNN for reproducibility.
• Introduced a deterministic methodology for RFNN training and evaluation.
• Employed Bayesian comparison to assess the significance of network modifications.

Main Results:

• Changing the network base had less impact than retraining with a different random seed.
• The simplified CNN architecture demonstrated performance comparable to the baseline RFNN.
• Energy normalization of filters positively impacted classification accuracy, even with random initialization.

Conclusions:

• Network architecture modifications, specifically base changes, may be less critical than the training process (seed selection).
• A simplified RFNN architecture can achieve performance similar to more complex baselines.
• Energy normalization is a beneficial technique for improving CNN classification accuracy.