Interpretable Artificial Intelligence through Locality Guided Neural Networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep learning models often feature complex, unorganized neurons, hindering human comprehension of their internal workings.
• Explainable Artificial Intelligence (XAI) research seeks to demystify 'black-box' AI methods.
• Current methods lack intuitive ways to visualize and understand neuron interactions within deep networks.

Purpose of the Study:

• To extend the Locality Guided Neural Network (LGNN) algorithm for improved explainability in deep learning.
• To organize neurons based on correlation, enabling the observation of interactions among neighboring neurons.
• To enhance the interpretability of deep learning models, particularly Convolutional Neural Networks (CNNs).

Main Methods:

• The study extends a previously developed algorithm, LGNN, which preserves locality between neighboring neurons during training.
• The approach is inspired by Self-Organizing Maps (SOMs) to enforce a local topology where neighboring neurons exhibit high correlation.
• LGNN is integrated into state-of-the-art CNN architectures like VGG and WRN for image classification tasks.

Main Results:

• Experiments on CIFAR100 and Imagenette datasets demonstrate that LGNN makes deep networks more interpretable both quantitatively and qualitatively.
• Visualizations reveal perceptible clusters of neuron activations corresponding to specific input classes.
• The method successfully increases the correlation between neighboring neurons in CNNs.

Conclusions:

• LGNN offers a method to organize neurons in deep learning, significantly enhancing model interpretability.
• The algorithm can be readily applied to existing CNNs, improving the understanding of their decision-making processes.
• This approach holds potential for broader applications beyond image processing in deep learning architectures.