Multi-Task Neural Network Training with Normalized Task Weights

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for training multi-task neural networks face challenges in balancing task weights effectively, leading to issues such as different loss scalings across tasks, resource imbalances, and computational inefficiencies, particularly in scenarios with highly imbalanced losses.

Innovation Solution

The method involves computing task weights by scaling individual losses to a constant using a normalization function, such as softmax, and applying a stop gradient operator to stabilize the weights, thereby normalizing them to ensure balanced learning across tasks.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If Uncertainty Weighting is used to dynamically adjust task weights, then task-specific optimal weights are achieved, but update inertia occurs and overfitting is prone

Engineering Contradiction:
Improvetask weight accuracyVSAvoidtraining stability
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent applies dynamics by making task weights adaptive rather than fixed. The uncertainty weights are dynamically adjusted during training based on the observed loss values, allowing the system to respond to changing task difficulties and data distributions. This is implemented through the uncertainty weighting mechanism that computes weights proportional to the inverse of loss variance, enabling automatic adaptation without manual intervention.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent changes the parameter of task weights from fixed or manually set values to dynamically computed uncertainty-based weights. By transforming the weight computation from static assignment to dynamic calculation based on loss statistics, the system achieves better weight accuracy while the normalization and smoothing techniques prevent instability and overfitting.

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If Scalarization with grid search is used to find optimal fixed weights, then superior performance is achieved, but computational cost becomes prohibitive

Engineering Contradiction:
Improveperformance accuracyVSAvoidcomputational time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent applies self-service by enabling the system to automatically determine optimal task weights without external intervention or exhaustive search. The uncertainty weighting mechanism autonomously computes appropriate weights based on the observed loss patterns during training, eliminating the need for manual grid search while achieving comparable or superior performance.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent replaces the mechanical grid search approach with a gradient-based uncertainty weighting mechanism. Instead of exhaustively searching through weight combinations, the system uses analytical solutions based on loss statistics to directly compute optimal weights, dramatically reducing computational time while maintaining accuracy.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

3Ease of manufacture

If Equal Weighting is used for all tasks, then implementation is simple, but drastically different loss scalings across tasks occur

Engineering Contradiction:
Improveimplementation simplicityVSAvoidloss scaling balance
Core Design Contradiction:
Ease of manufactureVSManufacturing precision

Solution Approach 1:

The patent changes the task weight parameter from uniform equal values to uncertainty-based computed values. This transformation maintains the simplicity of automatic weight assignment while correcting the loss scaling imbalances by adjusting weights according to observed task performance and loss characteristics.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent introduces dynamics by making weights adaptive rather than static. The uncertainty weighting mechanism continuously adjusts task weights based on observed loss values during training, allowing the system to automatically balance loss scalings across tasks with different magnitudes and difficulties.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP4604013A1Method for training a multi-task neural network
Publication Date: 2025.08.20 ROBERT BOSCH GMBH
  • EP4604013A1 patent drawingFigure 1a~2
  • EP4604013A1 patent drawingFigure 3a~3b
  • EP4604013A1 patent drawingFigure 4a

AI summary

Some embodiments are directed to training a multi-task neural network. The multi-task neural network is configured to receive an input and to produce multiple outputs. During training task weights are computed to weight individual losses. A normalization function is applied to normalize the task weights.