CNN Activation Map Visualization for Higher-Resolution Explainability
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current visualization methods for machine learning models, particularly convolutional neural networks, suffer from low resolution, making it difficult to interpret model behavior, especially for smaller objects or decisions based on finer details, due to the trade-off between resolution and semantic meaning.
Innovation Solution
Increase the input image resolution and adjust the resolution-independent convolutional layers of the neural network to generate higher resolution visualizations using methods like Grad-CAM, optionally adding pooling layers to match output sizes, without requiring additional training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If a convolutional layer close to the output of the CNN is selected for visualization, then the visualization is more semantically meaningful, but the resolution is reduced to very low levels (e.g., 7x7 pixels)
Solution Approach 1:
The patent applies another dimension by introducing a new spatial dimension through super-resolution techniques. Instead of accepting the low resolution (7x7 pixels) as a fixed constraint from the convolutional layer, the invention transforms the visualization by upsampling and refining it to higher resolutions (e.g., 224x224 pixels or higher), effectively adding a dimensional transformation that preserves semantic meaning while recovering fine-grained details.
Solution Approach 2:
The patent changes the resolution parameter of the visualization from low (7x7) to high (224x224 or higher) through super-resolution techniques. This parameter transformation allows the visualization to maintain its semantic accuracy from the deep convolutional layer while achieving the fine detail resolution needed for interpreting small objects and subtle features.
2Manufacturing precision
If the input image resolution is increased, then higher resolution visualizations can be generated, but the computational cost increases
Solution Approach 1:
The patent performs preliminary action by first obtaining the activation maps from the trained convolutional neural network at its native low resolution, then applying super-resolution techniques to upscale these maps. This two-stage approach avoids the computationally expensive alternative of running the entire CNN at high resolution, since only the activation map generation and subsequent upscaling are performed rather than full high-resolution forward propagation through all network layers.
Solution Approach 2:
The patent creates a copy of the activation information at a different resolution scale. Instead of processing the original high-resolution image through the entire network (which would be computationally intensive), the method copies the essential semantic information from the low-resolution activation maps and reconstructs high-resolution visualizations from these copies, significantly reducing computational requirements while maintaining visualization quality.
3Device complexity
If a single set of activation maps is used for classifying multiple different objects of different sizes, then the process is simplified, but the ability to detect smaller objects is compromised
Solution Approach 1:
The patent resolves this contradiction by applying another dimension through super-resolution upscaling. The single set of activation maps, regardless of object size in the original image, is transformed to high resolution, enabling small objects to be visualized with sufficient detail. This dimensional transformation allows the same activation maps to effectively represent both large and small objects at appropriate scales in the final visualization.
Solution Approach 2:
The patent applies local quality by enhancing different regions of the activation maps to different effective resolutions through the super-resolution process. Small objects in specific regions of the image receive enhanced local detail in the upscaling process, while large objects maintain their overall structure. This allows the single activation map set to provide appropriate detail levels for objects of varying sizes through localized refinement during the super-resolution transformation.
Data Source
AI summary
A method is provided for generating a visualization for explaining a behavior of a machine learning (ML) model. In the method, an image is input to the ML model for an inference operation. The input image has an increased resolution compared to an image resolution the ML model was intended to receive as an input. A resolution of a plurality of resolution-independent convolutional layers of the neural network are adjusted because of the increased resolution of the input image. A resolution-independent convolutional layer of the neural network is selected. The selected resolution-independent convolutional layer is used to generate a plurality of activation maps. The plurality of activation maps is used in a visualization method to show what features of the image were important for the ML model to derive an inference conclusion. The method may be implemented in a computer program having instructions executable by a processor.