Task-Specific Encoder Training for Explainable Multi-Estimator Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models used in automated driving and other applications have low explainability, making it difficult to analyze computation content and leading to potential malfunctions and decreased accuracy due to the complexity of computation parameters and local solutions.
Innovation Solution
A model generation apparatus that includes a data acquisition unit and a learning processing unit to train an encoder and multiple estimators, with specific training steps to ensure that feature amounts are exclusive to each estimation task, increasing explainability by minimizing correlations between different task-related information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If machine learning models are trained to execute estimation tasks, then the ability to perform regression or classification on unknown data is improved, but the explainability of computation content deteriorates
Solution Approach 1:
The patent segments the computation parameters into multiple groups, where each group corresponds to a specific estimation task. This segmentation allows the model to maintain high adaptability for executing multiple estimation tasks while improving explainability by associating specific parameter groups with specific tasks, making it easier to analyze and interpret the computation content for each task separately.
2Adaptability or versatility
If the number of computation parameters is increased to handle more complicated estimation tasks, then the estimation ability is improved, but the difficulty of reading and analyzing calculation content increases
Solution Approach 1:
The patent applies local quality by organizing computation parameters into distinct groups where each group has specific local characteristics related to particular estimation tasks. This allows the model to handle complicated estimation tasks with a large number of parameters while making analysis easier, as one can focus on and interpret the parameters relevant to specific tasks without being overwhelmed by the entire parameter set.
3Ease of manufacture
If machine learning models are trained with random initialization, then the training process is simplified, but the consistency of computation parameters across different models deteriorates
Solution Approach 1:
The patent applies preliminary action by establishing predefined groups of computation parameters before training begins. This preliminary organization of parameters into task-specific groups provides a structured framework that guides the training process, maintaining simplicity while ensuring that parameters with the same function remain consistent across different models, thereby improving reproducibility and reliability.
Data Source
AI summary
A model generation apparatus according to one or more embodiments executes: a first training step of training an encoder, a first estimator, and a second estimator so that, when training data of each of learning data sets is given to the encoder, a result of the first estimator executing a first estimation task matches first correct answer data, and a result of the second estimator executing a second estimation task matches second correct answer data; and a second training step of training the encoder so that, for each of the learning data sets, a correlation between a first portion of feature amounts and the second correct answer data is lowered, and a correlation between a second portion of the feature amounts and the first correct answer data is lowered.