Hierarchical Classification Tree for Perception Model Error Weighting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning perception models, such as convolutional neural networks, do not adequately differentiate between classification errors, treating all mistakes as equally wrong, which can lead to catastrophic errors in applications like autonomous vehicles, where certain mistakes are more severe than others.
Innovation Solution
The implementation of a hierarchically informed training method that uses a hierarchical classification tree to weight classification scores based on the hierarchical relationships between classes, encouraging the model to make 'better' mistakes by penalizing errors that are more severe in terms of safety or consequence.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a perception model is trained to minimize overall classification error, then the general accuracy is improved, but the severity of catastrophic errors increases because all errors are treated equally
Solution Approach 1:
The patent applies local quality by differentiating the treatment of different types of classification errors. Instead of treating all errors uniformly, the loss function assigns different weights to errors based on their severity and hierarchical distance from the ground truth class. This allows the model to focus more heavily on avoiding catastrophic errors while maintaining overall accuracy.
Solution Approach 2:
The patent changes the parameter weighting in the loss function to reflect the hierarchical structure of classification errors. By introducing hierarchical weights that increase with the distance from the ground truth class, the training process transforms the error landscape to penalize catastrophic mistakes more severely, thereby improving reliability without sacrificing accuracy.
2Reliability
If a hierarchical classification tree is used to weight classification scores, then the severity of errors is reduced, but the computational complexity increases
Solution Approach 1:
The patent applies preliminary action by pre-computing and storing the hierarchical classification tree structure before training begins. The hierarchical relationships and corresponding weights are established in advance, allowing the model to efficiently access pre-defined error severity metrics during training without performing complex hierarchical computations in real-time.
Solution Approach 2:
The hierarchical classification tree serves as an intermediary structure that mediates between the raw classification scores and the final loss calculation. Rather than directly computing complex error metrics, the tree provides a structured framework that translates classification distances into weighted penalties, simplifying the overall computational process.
Data Source
AI summary
A perception model is trained to classify inputs in relation to a discrete set of leaf node classes. A hierarchical classification tree encodes hierarchical relationships between the leaf node classes. A training loss function is dependent on a classification score for a given training input a its ground truth leaf node class of the training input, but also classification scores for at least some others of the leaf node classes, with the classification scores of the other leaf node classes weighted in dependence on their hierarchical relationship to the ground truth leaf node class within the hierarchical classification tree.