Understanding neural network through neuron level visualization.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Neural networks are fundamental to modern AI.
• Understanding the internal workings of neural networks, particularly the role of individual neurons, remains a challenge.
• Existing methods for neural network interpretability often require model modifications or are limited in scope.

Purpose of the Study:

• To propose and demonstrate a novel method for visualizing the learning process of individual neurons within neural networks.
• To enhance the interpretability of neural network models by making learned features human-understandable.
• To analyze the working mechanisms of different neural network architectures through feature visualization.

Main Methods:

• Developing a visualization technique to extract and display features learned by each neuron.
• Applying the method to Fully Connected Networks (FCNs) and Convolutional Neural Networks (CNNs).
• Utilizing the backpropagation learning algorithm for training neural network models.

Main Results:

• Successfully visualized neuron-learned features in an understandable format.
• Demonstrated the method's effectiveness on image classification tasks.
• Gained insights into the functional roles of neurons in FCNs and CNNs.

Conclusions:

• The proposed neuron visualization method effectively enhances neural network interpretability.
• The technique is versatile, applicable to various neural network architectures without modification.
• This approach provides a valuable tool for understanding and debugging complex neural network models.