Hierarchical Multi-Label Metric Learning for Few-Shot Embedded Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network metric learning methods fail to explain the relationship between classes in an embedding space, limiting their effectiveness in classifying new classes on embedded devices with small data sets.
Innovation Solution
A method for deep metric learning per hierarchical step of multi-labels and few-shot inference that samples multiplets from a hierarchical data set, configures them into triplets or quadruplets, and distributes a loss-function-removed deep neural network model to embedded devices for effective classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing neural network metric learning is used, then classification between classes is achieved, but the relationship between classes cannot be explained
Solution Approach 1:
The patent segments the embedding space into multiple hierarchical levels, where each level captures relationships at different granularities. This allows the system to maintain both fine-grained classification accuracy and coarse-grained class relationship structure simultaneously, resolving the contradiction between precise classification and relationship preservation.
Solution Approach 2:
The patent introduces hierarchical dimensionality to the embedding space, organizing classes across multiple levels rather than a single flat dimension. This dimensional transformation enables the preservation of class relationships while maintaining classification precision, as the hierarchical structure encodes semantic relationships that flat embeddings cannot capture.
2Loss of information
If deep metric learning with hierarchical multi-labels is implemented, then meaningful Euclidean distances and class relationships are achieved, but computational complexity increases
Solution Approach 1:
The patent performs preliminary organization of training data into hierarchical multi-label structures before the actual metric learning process. By pre-structuring the data with hierarchical relationships, the system reduces the computational burden during training, as the hierarchical framework is already established and does not require real-time computation during the learning phase.
Solution Approach 2:
The patent replaces complex iterative optimization mechanisms with a more efficient approach by using hierarchical triplet/quadruplet sampling combined with standard metric learning loss functions. This substitution maintains the ability to learn meaningful hierarchical relationships while reducing computational complexity compared to specialized hierarchical optimization algorithms.
3Loss of information
If metric learning is performed to achieve meaningful distances, then class relationships are preserved, but data requirements increase
Solution Approach 1:
The patent segments the learning task into hierarchical levels, where each level learns relationships at a specific granularity. This segmentation allows the system to learn meaningful class relationships from smaller data subsets at each hierarchical level, rather than requiring all possible class combinations to be present in the training data simultaneously, thus reducing overall data requirements.
Solution Approach 2:
The patent creates a universal hierarchical embedding framework that can be applied across different domains and class hierarchies. The hierarchical structure itself serves as a transferable prior that reduces data requirements, as the framework can leverage hierarchical relationships even when specific class examples are limited, making the system more data-efficient.
Data Source
AI summary
A method for deep metric learning per hierarchical step of multi-labels and few-shot inference using the same is provided, which includes sampling a plurality of N multiplets from a learning data set that is expressed by hierarchical multi-labels; configuring the N multiplets into triplets or quadruplets so as to correspond to the number of classes in each hierarchy; performing a metric learning per hierarchical step for a deep neural network model based on the configured triplets and quadruplets; and distributing, to an embedded device, the deep neural network model from which a loss function is removed as the metric learning is completed.