Multitasking Neural Network Segmentation for Independent Task Adaptation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Multitasking artificial neural networks face challenges in task-specific adaptations due to shared functions and intermediate calculations leading to dependencies between tasks, complicating validation and requiring re-validation of all tasks, even for adaptations intended for a single task.

Innovation Solution

A method for training a multitasking ANN with a first path for shared cross-task parameters and a second path for task-specific parameters, allowing independent adaptation of individual tasks without affecting others, utilizing a directed graph structure with cross-task and task-specific layers.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Use of energy by moving object

If shared functions and intermediate calculations are used in multitasking ANN, then computing resources such as computing time, storage space, memory bandwidth are saved, but dependencies between tasks are created complicating task-specific adaptations and requiring re-validation of all tasks

Engineering Contradiction:
Improvecomputing resourcesVSAvoidtask dependencies
Core Design Contradiction:
Use of energy by moving objectVSDevice complexity

Solution Approach 1:

The patent segments the ANN into task-specific sections that are independent from each other, while sharing common intermediate layers. This segmentation allows individual tasks to be adapted without affecting others, resolving the dependency issue while maintaining resource efficiency through shared common layers.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces a new structural dimension by adding task-specific sections beyond the shared intermediate layers. This dimensional extension allows tasks to be differentiated and adapted independently while still benefiting from the shared computational infrastructure, thus resolving the contradiction between resource efficiency and adaptability.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Adaptability or versatility

If adaptation is applied to a single task in multitasking ANN, then task-specific optimization is achieved, but all tasks must be re-validated due to shared functions

Engineering Contradiction:
Improvetask-specific adaptationVSAvoidvalidation time
Core Design Contradiction:
Adaptability or versatilityVSLoss of time

Solution Approach 1:

By segmenting the network into independent task-specific sections, the patent enables localized adaptation to single tasks without propagating changes to other tasks. This eliminates the need for re-validation of all tasks, significantly reducing validation time while maintaining adaptability.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent extracts task-specific functionality into separate sections that are independent from the shared common layers. This extraction allows adaptation to be applied to individual tasks without affecting the shared infrastructure or requiring re-validation of other tasks, thus reducing validation time.

Inventive Principle:
Principle #2Taking out (Extraction)

3Ease of operation

If multiple parallel single-task-capable ANNs are used, then each task can be executed independently, but computing resources such as computing time, storage space, memory bandwidth are consumed more heavily

Engineering Contradiction:
Improveindependent task executionVSAvoidcomputing resources
Core Design Contradiction:
Ease of operationVSUse of energy by moving object

Solution Approach 1:

The patent merges multiple single-task ANNs into a single multitasking ANN by sharing common intermediate layers and computational infrastructure. This merging maintains the ability to execute tasks independently through task-specific sections while significantly reducing computing resource consumption compared to running separate ANNs for each task.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The patent creates a universal multitasking ANN architecture where a single model can perform multiple tasks by sharing common computational layers. This multi-functionality allows the system to execute different tasks independently while consolidating resource usage, thus reducing overall computing resource consumption compared to dedicated single-task models.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP3918528B1Method for training and operating an artificial neural network capable of multitasking, artificial neural network capable of multitasking and apparatus
Publication Date: 2025.06.25 ROBERT BOSCH GMBH
  • EP3918528B1 patent drawingFigure 1~2
  • EP3918528B1 patent drawingFigure 3~4
  • EP3918528B1 patent drawingFigure 5

AI summary

The invention relates to an improved option for using a multitasking artificial neural network (KNN). In particular, the invention proposes a method for training a multitasking KNN (110). According to the invention, a first path (P1) for a first information flow through the KNN (110) is provided, wherein the first path (P1) couples an input layer (120) of the KNN (110) to at least one cross-task intermediate layer (130) of the KNN (110), which is common for a plurality of differing tasks of the KNN (110), and the first path (P1) couples the at least one cross-task intermediate layer to a respective task-specific KNN section (140) from the plurality of differing tasks (A, B). Furthermore, first training data for training cross-task parameters, which are common to the plurality of differing tasks of the KNN (110), is supplied via the input layer (120) and the first path (P1). In addition, at least one task-specific, second path (P2) for a second information flow, which is different from the first information flow, through the KNN (110) is provided, wherein the second path (P2) couples the input layer (120) of the KNN (110) to only one part of the task-specific KNN sections (140) from the plurality of differing tasks, and second training data for training task-specific parameters is supplied via the second path (P2).