Vision Transformer Pretraining With Token Sparsification

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing vision transformer training methods require large-scale datasets and are inefficient in self-supervised learning, particularly when token sparsification is applied to contrastive language-image pretraining, leading to data processing overhead and image-text misalignment.

Innovation Solution

A knowledge distillation framework that aligns image-text feature representations between a teacher and student encoder, incorporating token sparsification to accelerate training with uncurated datasets, reducing data processing overhead and improving image-text matching efficiency.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If global self-attention is used for vision transformer training, then vision task performance is improved, but data processing overhead increases excessively

Engineering Contradiction:
Improvevision task performanceVSAvoiddata processing efficiency
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The patent segments the full attention mechanism into sparse attention patterns by selectively masking out non-essential token interactions. This divides the computational workload from global O(N^2) attention to a subset of essential token pairs, maintaining performance while reducing processing overhead

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies partial attention by computing attention only for a subset of token pairs rather than all possible pairs. This partial computation approach maintains sufficient performance for vision tasks while significantly reducing the excessive data processing overhead of complete global self-attention

Inventive Principle:
Principle #16Partial or excessive action

2Productivity

If token sparsification is applied to accelerate pretraining, then training speed is improved, but image-text misalignment occurs in contrastive language-image pretraining

Engineering Contradiction:
Improvetraining speedVSAvoidimage-text alignment accuracy
Core Design Contradiction:
ProductivityVSManufacturing precision

Solution Approach 1:

The patent dynamically adjusts sparsification parameters and attention masking patterns based on the specific contrastive learning objectives. By changing the parameters of which tokens are sparsified and how attention is distributed, the method maintains image-text alignment accuracy while achieving training speed acceleration

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent incorporates feedback mechanisms that monitor alignment quality during sparse attention training and adjust the sparsification strategy accordingly. This feedback loop ensures that token sparsification does not degrade image-text alignment in contrastive language-image pretraining

Inventive Principle:
Principle #23Feedback

3Reliability

If large-scale datasets are used for complete global self-attention training, then model capability is improved, but training time and computational resources increase

Engineering Contradiction:
Improvemodel capabilityVSAvoidtraining time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent segments the training process into phases where sparse attention patterns are learned first, then progressively refined. This segmentation allows the model to capture essential features from large-scale datasets without requiring complete global self-attention computation throughout entire training

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent computes attention partially for large-scale dataset processing by focusing computational resources on the most informative token interactions. This partial computation approach maintains model capability development from large datasets while significantly reducing training time and computational resource requirements

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP4618036A1Method and system for pre-training vision transformer through knowledge distillation, and vision transformer pre-trained through same
Publication Date: 2025.09.17 LG MANAGEMENT DEV INST CO LTD
  • EP4618036A1 patent drawingFigure 1
  • EP4618036A1 patent drawingFigure 2
  • EP4618036A1 patent drawingFigure 3

AI summary

Disclosed is a method and system for pretraining vision transformers using large uncurated datasets in a self-supervised learning manner according to a knowledge distillation framework, thereby reducing data processing overhead and rapidly training simplified vision transformers.