Aggregated Digital Image Training for Neural Network Convergence
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional digital image search techniques are inefficient and prone to errors in domains with high variance, such as the fashion domain, due to reliance on manual attribute selection and iterative training processes, which consume significant computational resources and user interaction.
Innovation Solution
A two-stream machine learning architecture that leverages aggregated digital images to train neural networks, selecting positive and negative samples based on product IDs, allowing for faster convergence and reduced resource usage by processing multiple negative instances in a single pass.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If manual attribute selection is used for digital image search, then the system can handle high-variance domains like fashion, but it requires significant user interaction and is prone to errors
Solution Approach 1:
The system automatically selects attributes and generates annotations without requiring manual user input. The neural network performs self-service by extracting features and generating semantic annotations autonomously, eliminating the need for users to manually select attributes for each image.
Solution Approach 2:
The patent replaces the manual mechanical process of attribute selection with an automated neural network-based system. The mechanical interaction of users selecting attributes is substituted with computational processes that automatically extract and annotate image features.
2Reliability
If conventional iterative training processes are used, then the system can train neural networks for image search, but it consumes significant computational resources and time
Solution Approach 1:
The system performs preliminary actions by pre-processing images to generate aggregated representations before the main training process. The neural network is pre-trained on aggregated image data, which prepares it for faster and more efficient fine-tuning on specific datasets, reducing overall computational requirements.
Solution Approach 2:
Multiple images are merged into a single aggregated image representation that captures essential features across the dataset. This merging reduces the dimensionality and complexity of training data, allowing the neural network to learn from consolidated information rather than processing each individual image separately.
3Measurement precision
If multiple negative instances are processed in separate passes, then the training is thorough, but the computational time and resource usage increase
Solution Approach 1:
Multiple negative instances are merged into a single aggregated image that represents the combined information of all negative samples. This allows the neural network to process multiple negative instances simultaneously in a single pass rather than sequentially, reducing training time while maintaining training completeness.
Solution Approach 2:
The patent transforms the processing from a sequential temporal dimension (multiple passes) to a spatial dimension (single pass with aggregated data). By aggregating multiple negative instances into one composite representation, the system processes them in parallel within a single training pass, effectively adding a spatial consolidation dimension to reduce temporal processing requirements.
Data Source
AI summary
Digital image search training techniques and machine-learning architectures are described. In one example, a query digital image is received by service provider system, which is then used to select at least one positive sample digital image, e.g., having a same product ID. A plurality of negative sample digital images is also selected by the service provider system based on the query digital image, e.g., having different product IDs. The at least one positive sample digital image and the plurality of negative samples are then aggregated by the service provider system into a single aggregated digital image. At least one neural network is then trained by the service provider system using a loss function based on a feature comparison between the query digital image and samples from the aggregated digital image in a single pass.