Embedding Classification Training with Contrastive and Angular Losses
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural networks employed in vehicles for classification tasks often output erroneous decisions despite extensive training, necessitating improved accuracy.
Innovation Solution
A method involving the calculation of multiple types of losses, including contrastive learning and angular related losses, is applied to train neural networks by enhancing the representation of similar classes and sub-classes, using techniques like SWAV and ArcFace, to improve classification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If neural networks are extensively trained for classification tasks, then training time and computational resources are increased, but classification accuracy remains insufficient and erroneous decisions occur
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 optimization, and consolidation loss for final performance tuning. This segmented approach allows each phase to focus on specific aspects of accuracy improvement without requiring exhaustive training from scratch.
Solution Approach 2:
The method applies preliminary contrastive learning training before the main classification training. This preliminary action establishes strong initial representations of classes and sub-classes, which then serve as a foundation for subsequent angular related loss optimization, reducing the overall training time needed while improving accuracy.
2Measurement precision
If neural networks are trained to distinguish between classes and sub-classes, then representation capability is improved, but the network complexity and difficulty of training increase
Solution Approach 1:
The training is divided into distinct phases with specialized loss functions: contrastive learning loss handles the complex task of representing both classes and sub-classes, while angular related loss focuses on classification boundaries. This segmentation makes the overall complex training process more manageable and systematic.
Solution Approach 2:
The method changes the loss function parameters and training objectives across different phases. Starting with contrastive learning loss that emphasizes representation quality, then transitioning to angular related loss that emphasizes classification precision, and finally consolidation loss that optimizes overall performance. This parameter changes approach manages complexity by addressing different aspects sequentially.
Data Source
AI summary
A neural network training method for autonomous driving, the method includes (i) obtaining, by a computer device, features generated by a neural network and representing a first augmented image and at least a second augmented image, where the first augmented image and at least the second augmented image are different augmented image versions of a training image; (ii) determining, by the computer device, one or more an angular related losses based on the first augmented image and the second augmented image; (iii) determining a contrastive learning loss based on the first augmented image and the second augmented image; and (iv) updating the neural network based on the one or more angular related loss and on the contrastive learning loss.