Hash Generation Model Training for Stable Similarity Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional hash generation models suffer from poor performance and low accuracy due to oscillating optimization results during training, leading to inefficient use of hardware resources.
Innovation Solution
A method that considers both label similarity and hash similarity between sample object information to determine first and second loss values, training a hash generation model to generate accurate hash codes by iteratively adjusting these values.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If batch training is used for hash generation model, then training efficiency is improved, but optimization results oscillate and performance deteriorates
Solution Approach 1:
The patent implements feedback mechanisms by calculating loss values that compare hash similarity degrees against label similarity degrees. The model receives feedback through gradient descent optimization, adjusting parameters based on the difference between expected (label-based) and actual (hash-based) similarities. This feedback loop stabilizes optimization by continuously guiding the model toward consistent performance across batches.
Solution Approach 2:
The patent performs preliminary actions by pre-calculating label similarity degrees between all sample pairs before training begins. These pre-computed similarity values serve as stable reference targets during batch training, allowing the model to optimize hash generation against fixed ground-truth relationships rather than oscillating between varying batch compositions.
2Ease of manufacture
If conventional loss function is used, then model training is simple, but hash code generation accuracy is low
Solution Approach 1:
The patent transforms the loss function parameters from conventional single-value losses to multi-dimensional loss components. The new loss function incorporates label similarity degrees and hash similarity degrees as distinct parameters, allowing the model to simultaneously optimize for both semantic correctness (label-based) and hash code quality (similarity-based). This parameter expansion improves accuracy while maintaining computational feasibility through efficient similarity pre-computation.
Data Source
AI summary
This application involves an object information processing method, including obtaining a sample set, the sample set comprising sample object information; obtaining a degree of label similarity between labels corresponding to any two pieces of sample object information in the sample set, and obtaining a degree of hash similarity; determining a difference between the degree of label similarity and the degree of hash similarity, and obtaining a first loss value based on the difference; determining, for each piece of sample object information in the sample set, a first degree of hash similarity between the sample object information and similar sample object information; determining a second degree of hash similarity; determining a second loss value according to the first degree of hash similarity and the second degree of hash similarity; and training the to-be-trained hash generation model according to the first loss value and the second loss value.