Spatial Transformer Modules for Learning Image Transformation Invariance

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing neural networks struggle to efficiently learn invariance to transformations such as translation, scale, and rotation without additional training supervision or modifications.

Innovation Solution

Incorporating spatial transformer modules into neural networks that perform spatial transformations using a localisation subnetwork to generate transformation parameters, allowing the network to learn invariance through backpropagation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If traditional neural networks are used without spatial transformer modules, then the network structure remains simple, but the network cannot efficiently learn invariance to transformations such as translation, scale, and rotation

Engineering Contradiction:
Improveinvariance to transformationsVSAvoidnetwork structure
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The spatial transformer module acts as an intermediary component between the input layer and subsequent processing layers. It receives input features, generates spatial transformation parameters through a localisation subnetwork, applies spatial transformations (translation, scale, rotation), and outputs transformed features. This mediator enables the network to learn transformation invariance without requiring complex modifications to the entire network architecture.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The spatial transformer module is segmented into distinct functional components: a localisation subnetwork that generates transformation parameters, and a spatial transformation application component that applies these parameters. This segmentation allows each component to specialize in specific tasks while maintaining overall network simplicity and trainability through backpropagation.

Inventive Principle:
Principle #1Segmentation

2Adaptability or versatility

If additional training supervision or modifications are added to learn transformation invariance, then transformation invariance can be learned, but training complexity and computational overhead increase

Engineering Contradiction:
Improvetransformation invariance learningVSAvoidtraining time
Core Design Contradiction:
Adaptability or versatilityVSLoss of time

Solution Approach 1:

The spatial transformer module enables automatic feedback-based learning of transformation invariance. The localisation subnetwork generates transformation parameters that are applied to the input, and the transformed output is processed through subsequent layers. During backpropagation, gradients flow back through the entire pipeline, automatically adjusting the localisation subnetwork parameters to minimize loss. This eliminates the need for additional explicit supervision signals for transformation invariance while maintaining training efficiency.

Inventive Principle:
Principle #23Feedback

3Productivity

If spatial transformer modules are incorporated into neural networks, then training efficiency and classification accuracy improve, but the computational complexity of each forward pass increases

Engineering Contradiction:
Improveclassification accuracyVSAvoidcomputational complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The spatial transformer module dynamically changes parameters (transformation parameters) based on the input content rather than applying fixed transformations. The localisation subnetwork generates input-specific transformation parameters that adapt to each image's characteristics. This parameter adaptation enables the network to achieve state-of-the-art classification accuracy by focusing computational resources on relevant spatial transformations for each specific input, rather than processing all possible transformations uniformly.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3262569B1Spatial transformer modules
Publication Date: 2026.03.18 DEEPMIND TECH LTD
  • EP3262569B1 patent drawingFigure 1
  • EP3262569B1 patent drawingFigure 2
  • EP3262569B1 patent drawingFigure 3

AI summary

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for processing inputs using an image processing neural network system that includes a spatial transformer module. One of the methods includes receiving an input feature map derived from the one or more input images, and applying a spatial transformation to the input feature map to generate a transformed feature map, comprising: processing the input feature map to generate spatial transformation parameters for the spatial transformation, and sampling from the input feature map in accordance with the spatial transformation parameters to generate the transformed feature map.