Classification Model Calibration for Reliable Prediction Confidence

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing machine learning models, particularly neural networks, suffer from overfitting and overconfident predictions due to the imbalance between the number of model parameters and training samples, leading to unreliable classification decisions, especially in applications requiring accurate predictive confidence, such as sensor fusion and autonomous systems.

Innovation Solution

A calibration module is trained using output logit vectors from a trained classification model without retraining or adjusting its parameters, incorporating a finetuning submodule and optionally a binning submodule to adjust prediction probabilities, maximizing mutual information and using a modulation term based on prediction difficulty to improve calibration efficiency.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Extent of automation

If a trained classification model is used for classification tasks, then classification decisions can be made automatically, but the model produces overconfident predictions with inaccurate predictive confidence

Engineering Contradiction:
Improveautomatic classification decisionVSAvoidpredictive confidence accuracy
Core Design Contradiction:
Extent of automationVSReliability

Solution Approach 1:

A calibration module is introduced as an intermediary component between the trained classification model and the final prediction output. This calibration module adjusts the prediction probabilities to better reflect true confidence levels, resolving the overconfidence issue while preserving the automatic classification capability. The calibration module receives logits from the classification model and outputs calibrated probabilities that accurately represent predictive confidence.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If model parameters are adjusted to improve calibration, then predictive confidence accuracy improves, but the complexity of the solution increases significantly

Engineering Contradiction:
Improvepredictive confidence accuracyVSAvoidcalibration solution complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The calibration module is trained in advance on a calibration dataset to learn the relationship between raw prediction probabilities and true confidence levels. This preliminary training phase allows the module to be deployed with pre-computed calibration parameters, avoiding the need for complex real-time calibration computations during actual classification tasks. The calibration parameters are computed once and then applied repeatedly without additional computational overhead.

Inventive Principle:
Principle #10Preliminary action

3Measurement precision

If a large number of validation samples are used for calibration evaluation, then calibration accuracy improves, but the time and resources required increase significantly

Engineering Contradiction:
Improvecalibration evaluation accuracyVSAvoidcalibration evaluation time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The calibration module is trained on a relatively small calibration dataset, which is sufficient to achieve good calibration performance. The patent demonstrates that full calibration can be achieved without requiring exhaustive validation samples, as the calibration module learns the essential confidence-calibration relationship from a manageable dataset. This approach avoids the need for time-consuming evaluation with large numbers of validation samples while still achieving accurate calibration.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP3913538B1Classification model calibration
Publication Date: 2025.12.24 ROBERT BOSCH GMBH
  • EP3913538B1 patent drawingFigure 1~2
  • EP3913538B1 patent drawingFigure 3
  • EP3913538B1 patent drawingFigure 4

AI summary

A computer-implemented method (500) of calibrating a trained classification model is described. The trained classification model is trained to classify input samples according to a plurality of classes and to provide associated prediction probabilities, and comprises a plurality of hidden layers and at least one activation layer. The method comprises accessing (510) the trained classification model and accessing (510) a plurality of validation samples, each validation sample having a ground-truth label, the ground-truth label indicating a ground-truth class. The method further comprises applying (520) the trained classification model to the plurality of validation samples, obtaining (530), for each validation sample, an output logit vector from a layer of the trained classification model preceding a last activation layer, and training (540) a calibration module. The calibration module is trained to adjust prediction probabilities, the prediction probabilities being derived from the output logit vectors. The calibration module comprises at least one of a finetuning submodule for adjusting the prediction probabilities by finetuning the output logit vector and a binning submodule for adjusting the prediction probabilities by binning the output logit vector. The method further comprises appending (570) the trained calibration module to the trained classification model to obtain a calibrated classification model.