Feature Map Swapblock Augmentation for Neural Network Generalization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep neural networks face challenges in generalization due to over-fitting, particularly in handling object deformations not covered in the training dataset, which existing data augmentation methods struggle to address effectively.
Innovation Solution
A feature map-level data augmentation method, referred to as the swapblock algorithm, which exchanges values between randomly sampled units in the feature map, introducing variations in object deformations without altering the semantic meaning, implemented through a plug-in module called the E-module for both general and few-shot classification tasks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If image-level data augmentation methods are used to handle object deformations, then the training dataset can cover more deformation cases, but these methods are difficult to achieve and cannot effectively cover complicated real-world deformation cases
Solution Approach 1:
The patent transitions from image-level data augmentation to feature map-level augmentation, moving to a different dimension (from pixel space to feature space) where deformations can be implicitly modeled through value perturbations and unit exchanges, achieving both comprehensive coverage and ease of implementation
Solution Approach 2:
The patent introduces feature maps as an intermediary between input images and network parameters. By performing augmentation operations on feature maps rather than directly on images, the system can indirectly handle complex deformations through simpler operations like value perturbations and unit exchanges in the feature space
2Adaptability or versatility
If deep neural networks are trained with large number of parameters, then the network can learn complex patterns, but the network easily over-fits to the training dataset and poor generalization occurs
Solution Approach 1:
The patent applies data augmentation in advance during the training phase by introducing variations at the feature map level. This preliminary action exposes the network to diverse deformation patterns before actual testing, preventing over-fitting and improving generalization to unseen deformation cases
Solution Approach 2:
The patent changes the parameters of training data by introducing value perturbations and exchanging units in feature maps. This transforms the training data parameters to include implicit deformation variations, enabling the network to learn robust representations that generalize better to test data
Data Source
AI summary
Disclosed herein is a methodology implementing feature map-level data augmentation in a feature map. Two or more units in the feature map are selected and the values of locations in the two or more units are swapped among the two or more units. Value perturbations applied around local units in the feature map implicitly lead to an unused data augmentation at the image level.