CNN Kernel Labeling via Logic Program Extraction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for explaining convolutional neural networks (CNNs) lack global insights into feature interactions and fail to provide meaningful labels for kernels, leading to inefficient and computationally expensive processes for assigning symbolic labels.

Innovation Solution

A method that extracts a logic program from CNNs, uses feature-labeled image datasets to correlate kernel activations with features, and assigns labels based on highest correlation, reducing computational resource demand by processing support images only once per kernel.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Loss of information

If existing methods are used to explain CNN behavior through visualization, then local insights into individual samples are provided, but global insights into feature interactions and model-wide behavior are lacking

Engineering Contradiction:
Improveglobal insights into feature interactionsVSAvoidexplanation methodology complexity
Core Design Contradiction:
Loss of informationVSDevice complexity

Solution Approach 1:

The patent segments the CNN model into individual kernels and analyzes each kernel's contribution separately. By extracting and quantizing kernel outputs individually, then aggregating them to form logic programs, the method provides both local (individual kernel) and global (model-wide) insights without requiring complex visualization of entire feature interactions.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces logic programs as an intermediary representation between the CNN's internal kernel operations and human-understandable explanations. The logic programs serve as a mediator that translates complex kernel activations and feature interactions into symbolic rules that capture global model behavior while remaining interpretable.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Difficulty of detecting and measuring

If sophisticated visualization methods are used to backpropagate signals through weights and activations, then decomposition of model inner workings is achieved, but insight into interaction between features represented by other kernels is still lacking

Engineering Contradiction:
Improveinsight into feature interactionsVSAvoidcomputational resources
Core Design Contradiction:
Difficulty of detecting and measuringVSUse of energy by moving object

Solution Approach 1:

The patent extracts individual kernel outputs from the CNN and quantizes them to discrete values. By taking out each kernel's contribution separately and representing it symbolically in logic programs, the method enables analysis of feature interactions without requiring computationally intensive backpropagation through the entire network.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent changes the parameter representation of kernel outputs from continuous activation values to quantized discrete symbols. This parameter transformation enables efficient symbolic manipulation and interaction analysis in logic programs, avoiding the computational burden of working with continuous high-dimensional activation maps.

Inventive Principle:
Principle #35Parameter changes

3Loss of information

If methods are used to describe CNN classification decisions in graphs or trees, then relationships between different features are described, but negated symbols cannot be expressed

Engineering Contradiction:
Improveexpression of negated symbolsVSAvoidsymbolic representation capability
Core Design Contradiction:
Loss of informationVSDevice complexity

Solution Approach 1:

The patent uses logic programs as a universal symbolic representation that can express both positive and negative feature conditions within the same framework. The logic programming language naturally supports negation through standard logical operators, providing a multi-functional system that handles various types of feature relationships (positive, negative, interactions) uniformly.

Inventive Principle:
Principle #6Universality (Multi-functionality)

4Adaptability or versatility

If kernels are trained to be interpretable using proposed methods, then symbolic concepts are associated with kernels, but each kernel can belong to only one class and specific training methods are assumed

Engineering Contradiction:
Improvekernel class association flexibilityVSAvoidtraining method flexibility
Core Design Contradiction:
Adaptability or versatilityVSEase of manufacture

Solution Approach 1:

The patent makes the kernel-class association dynamic by extracting logic programs from already-trained CNNs rather than enforcing static one-to-one mappings during training. The quantization and logic program extraction process naturally handles kernels that may contribute to multiple classes, allowing the system to adapt to different CNN architectures and training methods without requiring retraining.

Inventive Principle:
Principle #15Dynamics

5Adaptability or versatility

If a prototype layer is learned to represent inputs in terms of similar components, then inputs are explained using specific kernels, but this assumes specific training methods and layer types

Engineering Contradiction:
Improveexplanation applicability to any CNNVSAvoidarchitecture constraints
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent creates a universal explanation method that works with any CNN architecture by extracting logic programs from the existing kernel structure rather than requiring specific prototype layers. The quantization and logic program extraction process is architecture-agnostic, making the explanation system versatile and applicable to any convolutional neural network without imposing architectural constraints.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP4131178A1Image classification method and apparatus, and method and apparatus for improving training of an image classifier
Publication Date: 2023.02.08 FUJITSU LTD
  • EP4131178A1 patent drawingFigure 1
  • EP4131178A1 patent drawingFigure 2
  • EP4131178A1 patent drawingFigure 3

AI summary

An image classification method comprises: extracting a logic program from a CNN, trained to classify features in images, which is a symbolic approximation of outputs of kernels at an extraction layer of the CNN; deriving kernel-based classification rules; forward-propagating pairs of feature-labeled images through the logic program to obtain kernel activations at the extraction layer for features in the images, where the scene in one of the pair contains a particular feature and the other is of the same scene without the feature; and calculating the correlation between each kernel in the logic program and each feature in the feature-labeled images using the kernel activations and the features associated with the feature-labeled images, assigning to each kernel in the logic program the label of the feature with which the kernel has the highest correlation, and applying the assigned kernel labels to the kernels in the rules to obtain kernel-labeled rules.