Image Retrieval Network Training with Randomized Multi-Learner Branches
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing multi-branch neural network models for image retrieval suffer from information errors accumulating across branches, leading to performance loss, network over-fitting, and poor attribute mapping due to branch sensitivity differences, affecting the entire model's performance.
Innovation Solution
Implementing a neural network training method where multiple learners share a common feature extractor and utilize random weights reassigned per training image group to enhance branch independence, reduce error accumulation, and improve robustness by ensuring invariance across image groups.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If multiple branches are trained in sequence with shared feature extractor, then branch differentiation capability is improved, but information errors accumulate across branches causing performance loss
Solution Approach 1:
The patent segments the training process by introducing independent random weight matrices for each branch, separating the error propagation paths of different branches while maintaining shared feature extraction capabilities. This allows each branch to learn distinct attribute mappings without accumulating errors from sequential training.
Solution Approach 2:
The patent changes the parameter representation by introducing random weight matrices as additional parameters in the feature extraction process. These random weights are specifically designed to vary across branches and training iterations, enabling branch differentiation while preventing error accumulation through parameter diversity.
2Reliability
If random weights are reassigned per training image group, then branch independence is enhanced, but computational complexity increases
Solution Approach 1:
The patent applies preliminary action by pre-generating random weight matrices before training begins and reassigning them in a structured manner during training. This preliminary preparation reduces the computational burden during actual training iterations while maintaining the benefits of branch independence.
Solution Approach 2:
The random weight matrices serve multiple functions: they enable branch differentiation, prevent error accumulation, and maintain computational efficiency through structured reassignment. This multi-functionality reduces the need for separate mechanisms for each objective.
3Ease of manufacture
If serialization training is used for multi-branch network, then training process is simplified, but error accumulation and over-fitting occur
Solution Approach 1:
The patent introduces random weight matrices as intermediary elements between the shared feature extractor and branch-specific learners. These intermediaries decouple the error propagation paths while maintaining the simplicity of sequential training, preventing over-fitting and error accumulation without complicating the training process.
4Productivity
If branches share common feature extractor, then model efficiency is improved, but information errors propagate across branches
Solution Approach 1:
The patent applies local quality by maintaining uniform random weight initialization across different spatial locations and branches of the network. This localized consistency in weight initialization prevents error propagation while preserving the efficiency benefits of shared feature extraction, as each location/branch processes information with identical statistical properties.
Data Source
Figure 1
Figure 2
Figure 3~4
AI summary
Provided are a neural network training method and apparatus for image retrieval, and an electronic device. A neural network comprises: one feature extractor and a plurality of learners. The method comprises: for each training image group, inputting three images of the training image group into the feature extractor, and determining features of the three images (501); for each image in each training image group, respectively multiplying the features of the image by a random weight corresponding to each learner, so as to obtain weighted features corresponding to each learner (502); for each image in each training image group, inputting the weighted features of the image corresponding to each learner into the corresponding learner, and determining a plurality of feature vectors of the image (503); and adjusting parameters of the neural network on the basis of the plurality of feature vectors of each image in a plurality of training image groups (504). The influence of information differences between training data on a network is weakened.