Neural Network Training via Local Feature Overlap Weighting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Training deep neural networks for image analysis tasks, such as classification and semantic segmentation, requires a substantial amount of annotated data, which is time-intensive and costly, and existing self-supervised learning methods like SimSiam are suboptimal for these tasks.
Innovation Solution
A method for training neural networks involves determining feature maps from transformed images and calculating loss values based on the overlap between image parts, using a weighted sum of feature vectors to improve feature representation and reduce mode collapse, allowing the network to learn fine-grained similarities and improve performance on downstream tasks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If SimSiam neural network is used for self-supervised learning, then the amount of annotated training data is reduced, but the performance on downstream tasks such as classification and semantic segmentation deteriorates because the network is configured to determine only a global feature representation of an entire image
Solution Approach 1:
The patent divides the global feature representation into multiple local feature vectors, each corresponding to a specific region or part of the image. This segmentation allows the network to capture fine-grained local features while still performing self-supervised learning, thereby maintaining high performance on downstream tasks like classification and semantic segmentation even with reduced annotated data
Solution Approach 2:
The patent applies local quality by making different parts of the feature representation serve different functions. Local feature vectors capture region-specific characteristics, while the weighted sum aggregation combines these local features into a global representation. This differentiation enables the network to simultaneously benefit from both local detail and global context, improving downstream task performance
2Loss of time
If conventional self-supervised learning methods are used, then training cost and time are reduced, but the ability to learn fine-grained similarities and part-based representations deteriorates
Solution Approach 1:
The network architecture segments the image into multiple regions and generates separate feature vectors for each region. This segmentation enables the learning of fine-grained local features during self-supervised training, capturing part-based similarities that conventional global pooling methods miss, all while maintaining efficient training times
Solution Approach 2:
The patent introduces a new dimension to the feature representation by organizing features both locally (per region) and globally (aggregated). This multi-dimensional feature space allows the network to learn fine-grained similarities at the local level while still achieving efficient training through the global aggregation mechanism
3Device complexity
If global feature representation is used, then the neural network training is simplified, but the performance on tasks requiring part-based understanding such as semantic segmentation deteriorates
Solution Approach 1:
The patent implements segmentation by dividing the global feature extraction process into multiple local feature vector computations, one for each image region. This segmentation enables part-based understanding for semantic segmentation tasks while keeping the overall network configuration relatively simple through the use of standard convolutional layers and pooling operations
Solution Approach 2:
The patent merges local feature vectors through a weighted sum operation to create a global feature representation. This combining mechanism allows the network to maintain simple global processing while incorporating detailed local information, thereby achieving good performance on part-based tasks without significantly increasing configuration complexity
Data Source
AI summary
A computer-implemented method for training a neural network. The training includes: determining a first feature map by the neural network based on a first transformed image, the first transformed image being determined based on a first transformation of a training image; determining a second feature map by the neural network based on a second transformed image, the second transformed image being determined based on a second transformation of the training image; determining a first loss value characterizing a metric between a first feature vector of the first feature map and a weighted sum of second feature vectors of the second feature map, weights of the weighted sum being determined according to overlaps of a part of the training image characterized by the first feature vector with respect to parts of the training image characterized by the respective second feature vectors; training the neural network based on the first loss value.