Embedding Classification Training With Customized Prototypes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural networks in vehicles often output erroneous classification decisions despite extensive training, necessitating improved accuracy in classification tasks.
Innovation Solution
A training process that calculates multiple types of losses for neural networks based on different augmentations of sensed information units, incorporating contrastive learning and angular related losses, particularly using SWAV and ArcFace methods, to enhance classification accuracy in both class and sub-class domains.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If neural networks are extensively trained, then training time increases, but classification accuracy remains insufficient
Solution Approach 1:
The training process is segmented into multiple phases with different loss functions: contrastive learning loss for initial representation learning, angular related loss for classification accuracy, and customized prototype-based loss for sub-class differentiation. This segmentation allows the model to learn progressively at different levels of abstraction without requiring excessive training time for all aspects simultaneously.
Solution Approach 2:
The method performs preliminary contrastive learning to establish good representations of sensed information units before applying more specialized angular and prototype-based losses. This preliminary action creates a solid foundation that reduces the additional training time needed for achieving high classification accuracy in specific domains.
2Reliability
If neural networks are trained with standard methods, then training process is simple, but classification accuracy in sub-class domains is insufficient
Solution Approach 1:
Different loss functions are applied at different stages and for different purposes: contrastive learning loss for general representation quality, angular related loss for class-level discrimination, and customized prototype-based loss for sub-class level discrimination. This local quality approach tailors the training process to specific needs at different levels without requiring complete redesign of the entire training pipeline.
Solution Approach 2:
The training process dynamically switches between different loss functions based on the training stage and specific requirements. The system adapts the loss function applied during training, transitioning from contrastive learning to angular loss to prototype-based loss as needed, making the training process flexible and responsive to different accuracy requirements.
Data Source
AI summary
A method for training a neural network, the method includes obtaining first augmentation image features and second augmentation image features for a training image associated with a specified class; selecting, out of different sets of contrastive learning loss prototypes, a set of contrastive learning loss prototypes associated with the specified class; wherein the different sets of contrastive learning loss prototypes are associated with different classes; determining a contrastive learning loss on the first augmentation image features and the second augmentation image features, using the selected set of prototypes; and updating the neural network based on the determined contrastive learning loss.