Multi-Label Neural Network Confidence Calibration for Reliable Predictions

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Machine learning models often output overly confident predictions, leading to confusion in downstream systems, and existing loss functions like cross-entropy encourage models to be over-confident, especially when training data is limited.

Innovation Solution

Implement a modified focal loss function during training, followed by temperature scaling to calibrate confidence, and generate a threshold-to-accuracy mapping to selectively discard predictions based on calibrated confidence levels.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If standard loss functions (cross-entropy) are used during training, then the model achieves good accuracy, but the model becomes over-confident and outputs unreliable confidence scores

Engineering Contradiction:
Improveconfidence score reliabilityVSAvoidconfidence score accuracy
Core Design Contradiction:
ReliabilityVSMeasurement precision

Solution Approach 1:

The patent modifies the loss function by introducing a temperature parameter T that scales the confidence scores. By optimizing this temperature parameter during training, the model learns to output confidence scores that better reflect the true reliability of predictions, resolving the contradiction between accuracy and reliability

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent performs confidence calibration as a preliminary step before deploying the model for production. By pre-calibrating the confidence scores using a validation set and determining appropriate temperature parameters, the system ensures reliable confidence scores are output from the start without affecting the model's predictive accuracy

Inventive Principle:
Principle #10Preliminary action

2Reliability

If the model outputs high confidence scores for all predictions, then the model appears confident, but downstream systems experience confusion due to unreliable confidence information

Engineering Contradiction:
Improveconfidence score trustworthinessVSAvoiduseful confidence differentiation
Core Design Contradiction:
ReliabilityVSLoss of information

Solution Approach 1:

The patent makes the confidence scoring dynamic by introducing temperature scaling that can be adjusted based on the specific prediction and class. This allows the model to output varied confidence scores that dynamically reflect the actual reliability of each prediction, preventing the collapse to uniform high confidence scores while maintaining useful differentiation for downstream systems

Inventive Principle:
Principle #15Dynamics

3Reliability

If temperature scaling is applied to calibrate confidence, then confidence reliability improves, but additional training iterations are required

Engineering Contradiction:
Improveconfidence calibrationVSAvoidtraining time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent applies temperature scaling only to the confidence calibration phase rather than retraining the entire model. By freezing the model weights and only optimizing the temperature parameter on a validation set, the approach achieves confidence calibration with minimal additional training time, rather than performing excessive full model retraining

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP3879462B1Calibrating reliability of multi-label classification neural networks
Publication Date: 2026.02.04 SAP SE
  • EP3879462B1 patent drawingFigure 1
  • EP3879462B1 patent drawingFigure 2
  • EP3879462B1 patent drawingFigure 3

AI summary

Methods, systems, and computer-readable storage media for tuning behavior of a machine learning (ML) model by providing an alternative loss function used during training of a ML model, the alternative loss function enhancing reliability of the ML model, calibrating the confidence of the ML model after training, and reducing risk in downstream tasks by providing a mapping between the confidence of the ML model to the expected accuracy of the ML model.