Unified Video Classifier Training via Category Graph Relationships
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In large-scale video hosting services, the lack of pre-categorized videos for training classifier functions limits their accuracy, as supervised learning requires a large representative training set, which is inefficient to create due to human labeling constraints, and conventional algorithms fail to capture inter-category relationships.
Innovation Solution
A classifier training system that trains unified classifiers by combining outputs from multiple initial classifiers, incorporating inter-category relationships and using a category graph to integrate diverse training information from various media types, including videos, web pages, and audio items, to enhance training accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised learning is used to train classifier functions, then classification accuracy is improved, but the requirement for large labeled training sets increases human labeling workload
Solution Approach 1:
The system performs preliminary classification using unsupervised learning to pre-label videos before human review. This preliminary action creates an initial labeled dataset that can be refined by human labelers, significantly reducing the total human labeling time required while maintaining classification accuracy.
Solution Approach 2:
The patent replaces manual human labeling (mechanical process) with automated unsupervised learning algorithms. The system uses feature extraction and clustering algorithms to automatically generate initial labels, substituting the time-consuming human labeling process with computational processes that can process videos at scale without human intervention.
2Measurement precision
If more videos are manually labeled to create a larger training set, then classifier accuracy is improved, but the cost and time required for human labeling increases
Solution Approach 1:
The system enables the training data to serve itself by using the video features and relationships within the video library to automatically generate labels. The unsupervised learning algorithms exploit inherent patterns in the video data to create labels without requiring external human annotation for each video, allowing the system to bootstrap its own training process.
Solution Approach 2:
The system creates a multi-functional approach where the same video features extracted for content understanding are also used for automatic labeling. The feature extraction process serves dual purposes: both representing video content for classification and providing the basis for unsupervised label generation, eliminating the need for separate labeling processes.
3Device complexity
If conventional classification algorithms are used that assume category independence, then the classification process is simpler, but inter-category relationship information is lost
Solution Approach 1:
The patent introduces a category relationship graph as an intermediary structure that captures inter-category relationships. This graph serves as a mediator between the classification algorithm and the category hierarchy, allowing the system to incorporate relationship information without fundamentally complicating the core classification process. The graph structure encodes parent-child and sibling relationships that inform the classification decisions.
Solution Approach 2:
The system adds another dimension to the classification problem by incorporating the category relationship graph. Instead of treating classification as a flat, independent category assignment problem, the patent embeds the hierarchical and relational structure of categories as an additional dimensional constraint that guides the classification process, allowing relationships to be captured without excessive complexity.
4Reliability
If only manually labeled videos are used for training, then the training data quality is high, but the training set size remains too small for effective classification
Solution Approach 1:
The patent merges manually labeled videos with automatically labeled videos from the unsupervised learning process to create a combined training set. This combination leverages the high quality of manual labels for a smaller subset while incorporating the larger volume of automatically generated labels, achieving both adequate training set size and sufficient data quality through the aggregation of multiple labeling sources.
Data Source
AI summary
A classifier training system trains unified classifiers for categorizing videos representing different categories of a category graph. The unified classifiers unify the outputs of a number of separate initial classifiers trained from disparate subsets of a training set of media items. The training process takes into account the relationships that exist between the various categories of the category graph by relating scores associated with related categories, thus enhancing the accuracy of the unified classifiers.