Label Refinery for Dynamic Image Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning algorithms face challenges in image classification due to incomplete, ambiguous, and redundant labels, which affect the accuracy and generalization of models.
Innovation Solution
The introduction of a Label Refinery, an iterative process that refines ground truth labels by examining the entire dataset, generates soft, informative, collective, and dynamic labels, improving the accuracy of image classification models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional labeling methods are used for training data, then the labeling process is simple and fast, but the labels are incomplete, ambiguous, and redundant which reduces classification accuracy
Solution Approach 1:
The patent applies preliminary action by performing label refinement before the actual model training process. A refinement model is trained on the original dataset with initial labels, and its predictions are used to generate refined labels that are then used for subsequent model training. This preliminary label refinement step improves classification accuracy by addressing incomplete, ambiguous, and redundant labels before the main training occurs.
Solution Approach 2:
The patent introduces an intermediary refinement model that acts as a mediator between the original training data and the final classification model. This refinement model processes the original labels and generates improved labels, serving as an intermediate step that bridges the gap between simple traditional labeling and high-accuracy classification requirements.
2Measurement precision
If iterative label refinement is applied to improve label quality, then classification accuracy improves, but the processing time and computational resources increase
Solution Approach 1:
The patent applies partial action by implementing a controlled iterative refinement process where the refinement model is trained for a specific number of epochs (e.g., 90 epochs) rather than continuously. The process stops when convergence is achieved or a predetermined number of iterations is reached, balancing the improvement in top-1 accuracy with acceptable processing time constraints.
Solution Approach 2:
The patent employs periodic action through iterative epochs of training the refinement model. The model undergoes multiple training cycles (epochs) where it periodically processes the data, refines labels, and updates its predictions. This periodic iterative process allows the system to progressively improve label quality and classification accuracy while managing computational resources through structured training cycles.
Data Source
AI summary
Systems and methods are disclosed for training neural networks using labels for training data that are dynamically refined using neural networks and using these trained neural networks to perform detection and/or classification of one or more objects appearing in an image. Particular embodiments may generate a set of crops of images from a corpus of images, then apply a first neural network to the set of crops to obtain a set of respective outputs. A second neural network may then be trained using the set of crops as training examples. The set of respective outputs may be applied as labels for the set of crops.