Nested Vision Transformer Aggregation for Data-Efficient Classification

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Vision transformers require significant amounts of data for effective performance and lack inductive bias, leading to inefficiencies and worse performance on smaller datasets compared to CNNs.

Innovation Solution

A nested hierarchical architecture for vision transformers that aggregates feature representations using convolutional and pooling operations, maintaining self-attention range to improve accuracy and data efficiency while simplifying the architecture.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If vision transformers use hierarchical structures to improve performance, then accuracy is improved, but device complexity and data requirements increase significantly

Engineering Contradiction:
Improveimage classification accuracyVSAvoidmodel architecture complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent implements a nested hierarchical structure where multiple sets of transformers are organized in nested levels. Each set processes feature representations at a specific hierarchy level, with outputs fed into the next level. This nesting approach systematically builds complexity while maintaining modular organization, resolving the contradiction by providing structured accuracy improvement without chaotic complexity increase.

Inventive Principle:
Principle #7Nested doll (Nesting)

Solution Approach 2:

The vision transformer model is segmented into multiple distinct sets of transformers, each responsible for specific processing tasks at different hierarchical levels. This segmentation allows independent optimization of each set while maintaining overall system coherence, improving accuracy through specialized processing without requiring the entire system to become uniformly complex.

Inventive Principle:
Principle #1Segmentation

2Measurement precision

If vision transformers use hierarchical structures to improve accuracy, then performance is improved, but quantity of data required increases significantly

Engineering Contradiction:
Improveimage classification accuracyVSAvoidquantity of training data
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent changes key parameters of the transformer architecture, specifically the hierarchical organization into multiple sets with different configuration parameters. Each set operates with optimized parameters for its specific processing level, allowing the model to achieve high accuracy on smaller datasets by leveraging parameter-efficient hierarchical processing rather than requiring massive data volumes.

Inventive Principle:
Principle #35Parameter changes

3Measurement precision

If vision transformers maintain self-attention range to improve accuracy, then measurement precision is improved, but productivity and processing speed decrease

Engineering Contradiction:
Improvefeature representation accuracyVSAvoidprocessing speed
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The processing workload is segmented across multiple sets of transformers operating at different hierarchical levels. Each set handles a specific portion of the computation with optimized attention ranges appropriate to its level, maintaining accuracy where needed while improving overall processing throughput by distributing the computational burden.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The hierarchical structure enables periodic processing where each set of transformers operates in sequence through distinct processing stages. This periodic action allows the system to maintain comprehensive self-attention capabilities while organizing computations in efficient batches, balancing accuracy requirements with processing speed through structured temporal organization.

Inventive Principle:
Principle #19Periodic action

Data Source

PatentEP4348599B1Aggregating nested vision transformers
Publication Date: 2025.10.15 GOOGLE LLC
  • EP4348599B1 patent drawingFigure 1
  • EP4348599B1 patent drawingFigure 2
  • EP4348599B1 patent drawingFigure 3

AI summary

A method (400) includes receiving a series of image patches (142) of an image (140). The method includes generating, using a first set of transformers (242) of a vision transformer (V-T) model (202), a first set of higher order feature representations (244) based on the image patches and aggregating the first set of higher order feature representations into a second set of higher order feature representations (246) that is smaller than the first set. The method includes generating, using a second set of transformers (248) of the V-T model, a third set of higher order feature representations (250) based on the second set and aggregating the third set of higher order feature representations into a fourth set of higher order feature representations (252) that is smaller than the third set. The method includes generating, using the V-T model, an image classification (170) of the image based on the fourth set.