Capsule Network Discrimination Basis Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Capsule networks struggle to provide clear discrimination basis for class discrimination, as the underlying reasons for output class decisions remain unknown and difficult to interpret.
Innovation Solution
A method and apparatus for a capsule network that learns to reproduce correspondence between data elements and their pre-labels, acquiring intermediate data based on activation and pose, and calculating similarity between input data and learned data to determine discrimination basis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a capsule network is used for class discrimination, then classification accuracy is improved, but interpretability of the discrimination basis deteriorates
Solution Approach 1:
The patent introduces an intermediary explanation generation module that translates the internal capsule network representations (pose and activation) into human-interpretable discrimination bases. This mediator converts the abstract mathematical operations into meaningful explanations about which features contributed to the classification decision, thus preserving both accuracy and interpretability.
Solution Approach 2:
The patent replaces the opaque black-box mechanism of traditional capsule networks with a transparent explanation system. By substituting the unknown internal decision-making process with an explicit explanation generation mechanism that traces back to input features, the system maintains classification performance while providing interpretable discrimination bases.
2Loss of information
If intermediate data is acquired from capsule layers for explanation, then discrimination basis clarity is improved, but computational complexity increases
Solution Approach 1:
The patent extracts only the essential intermediate data (pose and activation values) from the capsule layers that are necessary for generating explanations, rather than processing all internal network states. This selective extraction provides sufficient discrimination basis clarity while minimizing the additional computational burden.
Solution Approach 2:
The patent implements a partial explanation approach where only the most relevant capsule layers and features are analyzed for generating discrimination bases. By performing partial action on the most informative subsets of data rather than exhaustive analysis of all network components, the system achieves adequate explanation clarity with reduced computational complexity.
Data Source
AI summary
A method for causing one or more processors to execute: performing learning of a first model of a capsule network type including one or more capsule layers each having one or more capsules to reproduce correspondence between a plurality of first data elements included in a first data set and a pre-label corresponding to each of the plurality of first data elements; and inputting the first data set into the learned first model and acquiring first intermediate data based on at least one of a first activation and a first pose included in the one or more capsules, for the one or more capsule layers.