Convolutional Neural Network Partial Annotation Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning approaches for multi-label image classification face challenges with partial annotations, where only some labels are known, and the annotation cost is high, leading to impractical labeling for humans, especially with increasing dimensions and categories.
Innovation Solution
A computational approach is introduced that uses a new classification loss relation and normalization function to train convolutional neural networks with partial annotations, leveraging graph neural networks for modeling interactions between categories and predicting missing labels through curriculum learning and Bayesian uncertainty strategies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If more training data with complete labels is collected to improve classification performance, then model accuracy improves, but annotation cost increases proportionally to the number of images and categories
Solution Approach 1:
The patent applies partial action by training models with only a subset of labels (partial annotations) rather than requiring complete labels for all categories. The system learns to predict missing labels through the architecture itself, eliminating the need for full annotation of every image-category combination and significantly reducing annotation costs while maintaining classification performance
Solution Approach 2:
The patent implements self-service by enabling the model to generate its own labels through the convolutional neural network architecture. The network uses its learned features and the provided partial labels to predict missing labels, effectively self-completing the annotation process without requiring human annotators to provide complete labels for all categories
2Adaptability or versatility
If the number of labels and categories increases to capture more information, then classification capability improves, but the complexity of labelling and annotation increases
Solution Approach 1:
The patent applies partial action by requiring only partial labels rather than complete labels for all categories. This allows the system to handle multiple categories and increase classification capability while avoiding the complexity of annotating every category for every image, as the model infers missing labels automatically
Solution Approach 2:
The patent replaces the mechanical manual annotation process with a computational system. Instead of requiring human annotators to manually label all categories, the convolutional neural network architecture performs the labeling computation automatically, substituting manual mechanical labelling with automated computational inference
3Measurement precision
If manual labelling is conducted to ensure accurate annotations, then label accuracy improves, but the time required for labelling increases
Solution Approach 1:
The patent implements self-service by having the model generate its own labels through the convolutional neural network architecture. The network uses its learned features and provided partial labels to predict missing labels, eliminating the need for time-consuming manual annotation of all categories and significantly reducing labelling time while maintaining accuracy
Solution Approach 2:
The patent replaces the mechanical manual labelling process with automated computational inference. The convolutional neural network performs label prediction through matrix operations and activation functions, substituting the time-consuming manual process with fast computational operations that maintain label accuracy
Data Source
AI summary
Effectively training machine learning systems with incomplete/partial labels is a practical, technical problem that solutions described herein attempt to overcome. In particular, an approach to modify loss functions on a proportionality basis is noted in some embodiments. In other embodiments, a graph neural network is provided to help identify correlations/causations as between categories. In another set of embodiments, a prediction approach is described to, based on originally provided labels, predict labels for unlabelled training samples such that the proportion of labelled labels relative to all labels is increased.