Neural Network Importance Scoring With Pairwise Loss Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional machine learning techniques face challenges in effectively training importance scoring models for autonomous vehicles to accurately assign importance scores to agents in their vicinity, particularly due to computational resource limitations and the difficulty in evaluating errors related to background positions with arbitrary importance scores.
Innovation Solution
The system trains a neural network to focus on locations with the highest ground truth importance scores by setting training scores for background positions to zero, thereby preventing loss computations from affecting the training process, and uses a pair-wise loss function to optimize the objective function, allowing for more accurate and efficient agent importance estimation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional machine learning techniques are used to train importance scoring models, then the model can process all locations in the input, but the training process is computationally expensive and time-consuming due to considering background positions with arbitrary importance scores
Solution Approach 1:
The patent extracts and removes background positions from the training process by setting their importance scores to zero. This selective extraction focuses computational resources only on agent positions that matter, eliminating the time-wasting computation on irrelevant background locations while preserving accurate importance scoring for actual agents
Solution Approach 2:
The patent applies different treatment to different locations: agent positions retain their importance scores for accurate modeling, while background positions are set to zero. This local differentiation optimizes training efficiency by concentrating computational effort where it is most needed - on actual agents rather than background areas
2Measurement precision
If conventional machine learning techniques compute loss for all locations including background positions, then comprehensive error evaluation is performed, but computational resources are wasted on positions with arbitrary importance scores
Solution Approach 1:
The patent extracts only the relevant subset of locations (agent positions) for loss computation by setting background position scores to zero. This extraction maintains accurate error evaluation for actual agents while eliminating wasted computational energy on background positions with arbitrary scores
3Reliability
If the model processes all locations in the input, then complete coverage is achieved, but the training efficiency decreases due to the large number of background positions
Solution Approach 1:
The patent extracts and isolates agent positions from background positions by setting the latter to zero. This extraction maintains reliable importance estimation for agents while dramatically improving training speed by reducing the number of positions that require computational processing
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training a neural network configured to receive a network input and to assign a respective score to each of a plurality of locations in the network input. In one aspect, a method includes obtaining a training input and a corresponding ground truth output; processing the training input to generate a training output; computing a loss for the training input, comprising: selecting a plurality of candidate locations; setting to zero the training scores for any location in the selected candidate locations that has a ground truth score below a threshold value; for each of a plurality of pairs of locations in the selected candidate locations: computing a pair-wise loss for the pair; and combining the pair-wise losses to compute the loss for the training input; and determining an update to the current values of the parameters.