An Interactive Visualization for Feature Localization in Deep Neural Networks.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Deep artificial neural networks, particularly deep convolutional neural networks (CNNs), excel at computer vision tasks but suffer from opacity, acting as 'black boxes'.
• Existing methods for visualizing neural network workings are crucial for understanding and improving these complex models.
• Interpreting the decision-making process of deep neural networks is essential for reliable AI applications.

Purpose of the Study:

• To introduce Interactive Feature Localization in Deep neural networks (IFeaLiD), a novel tool for visualizing convolutional neural network layers.
• To provide a method for interpreting how different layers of a CNN perceive image regions by analyzing feature vector similarities.
• To facilitate the design, inspection, and interpretation of deep neural networks in computer vision.

Main Methods:

• IFeaLiD interprets neural network layers as multivariate feature maps.
• It visualizes the similarity between feature vectors of individual pixels in a heatmap display.
• The tool utilizes GPU acceleration with WebGL 2 for real-time processing of high-resolution feature maps in a web browser.

Main Results:

• The similarity display reveals how input images are perceived by different CNN layers.
• It allows comparison of the perception of specific image regions against the rest of the image.
• Examples from four computer vision datasets using a pre-trained ResNet101 demonstrate the tool's capabilities.

Conclusions:

• IFeaLiD offers an effective and interactive approach to visualizing and understanding the internal representations of deep convolutional neural networks.
• The tool enhances the interpretability of 'black box' models, aiding in debugging and model development.
• IFeaLiD is open-source and accessible online, promoting wider adoption in the computer vision research community.