AI Classification Using Unlabeled Data and Class Descriptors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing classification techniques face challenges in handling large numbers of target classes with complex attributes, requiring large labeled datasets and often suffer from overfitting due to reliance on unrepresentative training data, especially when classes are hierarchically organized and differ by small details.
Innovation Solution
The method employs a combination of labeled and unlabeled data to learn a class-weights matrix, using feature processing algorithms to identify compatible and incompatible attributes, which are then used to train a classification model. This approach allows for iterative training and parallelization, reducing resource requirements and avoiding overfitting by incorporating information from both labeled and unlabeled data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep neural network based classification techniques are used, then classification accuracy can be improved, but large numbers of labeled data items are required which are hard to acquire when there are thousands of target classes
Solution Approach 1:
The patent segments the classification task into multiple stages: first training on a subset of target classes with labeled data, then progressively expanding to additional classes. This segmentation allows the system to achieve good classification accuracy without requiring labeled data for all thousands of target classes simultaneously, addressing the contradiction between accuracy and data quantity requirements
Solution Approach 2:
The patent performs preliminary action by pre-training the neural network on a carefully selected subset of target classes that have available labeled data before deploying to handle the full set of thousands of classes. This preliminary training establishes a solid foundation that can be extended to other classes without requiring complete labeled datasets for all classes upfront
2Productivity
If traditional classification techniques rely heavily on labeled training data, then model training can be completed, but the models suffer from unrepresentative training data and cannot generalize well to new data items with substantially different characteristics
Solution Approach 1:
The patent applies universality by designing a classification model trained on a subset of classes that can universally generalize to handle thousands of target classes beyond the training set. The model learns transferable features and patterns that make it applicable across different class domains, not just the specific classes used during training, thereby improving reliability and generalization capability
Solution Approach 2:
The patent uses copying by learning from labeled examples and applying those learned patterns to unlabeled or partially labeled data. The model copies the classification behavior learned from representative labeled data to make predictions on new data items with different characteristics, enabling generalization without requiring exhaustive labeled datasets
3Adaptability or versatility
If large training datasets are collected to improve classification performance, then more comprehensive coverage of target classes is achieved, but the difficulty of acquiring labeled data increases significantly when there are thousands of target classes
Solution Approach 1:
The patent segments the target classes into multiple groups or subsets, training the model on strategically selected subsets that provide diverse and representative coverage. This segmentation approach allows the system to achieve broad adaptability across thousands of classes without requiring comprehensive labeled data for all classes, thereby reducing data acquisition complexity while maintaining versatility
Solution Approach 2:
The patent applies partial action by training on a carefully selected portion of target classes that provides sufficient diversity and representativeness. Rather than attempting to label all thousands of classes exhaustively, the system uses partial coverage of strategically important classes to achieve good generalization performance, reducing the complexity of data acquisition while maintaining adaptability
Data Source
AI summary
At an artificial intelligence service, a first group of attributes of data items, which can be represented in the same encoding space as a second group of attributes of class descriptors, is identified. Feature data structures are generated that respectively represent (a) the first group of attributes of an unlabeled subset of the data items and (b) the second group of attributes of the class descriptors. In one or more training iterations, a class-weights data structure that meets a quality criterion is computed, where the computations of at least one iteration include generating an aggregate data structure from the feature data structures. The class-weights data structure is stored.