Object Classifier Neural Networks With Two-Phase Coarse-to-Fine Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles face challenges in accurately classifying objects using neural networks, as existing systems degrade in coarse-object classification performance when transitioning from coarse to fine-object classification tasks, leading to reduced accuracy and increased computational inefficiency.
Innovation Solution
A two-phase training method for object classification neural networks, where the first phase optimizes coarse-object classifications and the second phase refines fine-object classifications by freezing parameters of channel encoder subnetworks not involved in fine-object classification, ensuring minimal degradation of coarse-object classification capabilities and efficient training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the neural network is trained to perform fine-object classification, then the fine-object classification accuracy is improved, but the coarse-object classification performance degrades
Solution Approach 1:
The neural network is divided into multiple independent channel encoder subnetworks, each responsible for processing specific object categories. This segmentation allows the network to specialize in fine-object classification for certain categories while maintaining robust coarse-object classification capabilities across all categories, resolving the performance degradation issue.
Solution Approach 2:
Different portions of the neural network are assigned different functions: some channel encoder subnetworks are optimized for fine-object classification while others maintain general coarse-object classification capabilities. This local differentiation allows the system to achieve high fine-object accuracy without sacrificing overall coarse-object performance.
2Measurement precision
If all parameters are adjusted during fine-object classification training, then the fine-object classification is optimized, but the training computational overhead increases
Solution Approach 1:
The patent extracts and freezes the parameters of channel encoder subnetworks that are not involved in fine-object classification during the second training phase. This extraction approach allows training to focus only on the necessary subsets of parameters, significantly reducing computational overhead while maintaining fine-object classification accuracy.
Solution Approach 2:
Instead of adjusting all parameters during fine-object classification training, the system applies partial action by only adjusting parameters of channel encoder subnetworks relevant to the specific fine-object category being trained. This selective parameter adjustment reduces training computational overhead while achieving the desired optimization.
3Device complexity
If the neural network is designed for coarse-object classification only, then the system complexity is low, but the system cannot perform fine-object classification
Solution Approach 1:
The neural network is designed with multiple channel encoder subnetworks that can serve different functions: some are optimized for coarse-object classification while others are specialized for fine-object classification. This multi-functional design allows the system to perform both coarse and fine-object classification without requiring separate independent systems, balancing complexity and versatility.
Data Source
AI summary
Aspects of the subject matter disclosed herein include methods, systems, and other techniques for training, in a first phase, an object classifier neural network with a first set of training data, the first set of training data including a first plurality of training examples, each training example in the first set of training data being labeled with a coarse-object classification; and training, in a second phase after completion of the first phase, the object classifier neural network with a second set of training data, the second set of training data including a second plurality of training examples, each training example in the second set of training data being labeled with a fine-object classification.