Zero-Shot Learning with Multi-Scale Manifold Alignment

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing zero-shot learning methods suffer from decoupled correlation processes between visual and semantic features, which are sensitive to noise, especially in automated semantic attributes, limiting their performance in real-world applications.

Innovation Solution

A system using localized multi-scale graph transform for manifold alignment, aligning visual and semantic representations through Spectral Graph Wavelets (SGW) regularization, integrating the correlation process into a single unified step, and applying Tikhonov regularization to SGW coefficients for unsupervised learning.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Extent of automation

If automated semantic attributes are used for zero-shot learning, then the system can operate without human annotation and scale better, but the semantic representations become very noisy which limits performance

Engineering Contradiction:
Improveautomation of attribute generationVSAvoidnoise in semantic representations
Core Design Contradiction:
Extent of automationVSReliability

Solution Approach 1:

The patent introduces spectral graph wavelets as an intermediary mechanism between automated semantic attributes and visual features. The SGW transform acts as a mediator that processes noisy semantic representations through spectral filtering, separating signal from noise in the frequency domain and producing cleaned semantic features for accurate zero-shot learning.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent replaces traditional mechanical/algorithmic noise suppression methods with spectral graph wavelet analysis. Instead of using conventional denoising techniques, the system substitutes a frequency-domain spectral transformation that leverages the mathematical properties of wavelets to inherently filter noise while preserving semantic information.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

2Ease of manufacture

If the correlation process between visual and semantic features is decoupled into separate optimization steps, then each step can be optimized independently, but the overall process becomes complex and less effective at suppressing noise

Engineering Contradiction:
Improveseparate optimization stepsVSAvoidcomplexity of correlation process
Core Design Contradiction:
Ease of manufactureVSDevice complexity

Solution Approach 1:

The patent merges the correlation process between visual and semantic features into a single unified optimization step using spectral graph wavelet analysis. Instead of separate optimization steps, the system combines feature alignment and noise suppression into one integrated SGW-based manifold alignment process, simplifying the overall architecture while improving noise suppression effectiveness.

Inventive Principle:
Principle #5Merging (Combining)

3Adaptability or versatility

If k nearest neighbor parameter selection is used on the graph, then the method can classify unseen instances, but the performance becomes very sensitive to noise in the graph construction

Engineering Contradiction:
Improveclassification of unseen instancesVSAvoidsensitivity to noise
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The patent applies preliminary spectral graph wavelet transformation and manifold alignment before performing k-nearest neighbor classification. By pre-processing the data through SGW denoising and establishing a cleaned manifold structure, the system prepares the data in advance to be less sensitive to noise during the subsequent classification step, improving robustness of unseen instance recognition.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3563296B1Zero-shot learning using multi-scale manifold alignment
Publication Date: 2025.11.12 HRL LAB
  • EP3563296B1 patent drawingFigure 1
  • EP3563296B1 patent drawingFigure 2
  • EP3563296B1 patent drawingFigure 3

AI summary

Described is a system for recognition of unseen and untrained patterns. A graph is generated based on visual features from input data, the input data including labeled instances and unseen instances. Semantic representations of the input data are assigned as graph signals based on the visual features. The semantic representations are aligned with visual representations of the input data using a reguiarizalicm method, applied, directly in a spectral graph wavelets (SOW) domain. The semantic representations are then used to generate labels for the unseen instances. The unseen instances may represent unknown conditions tor an autonomous vehicle.