Active Learning Neural Network for Image Labeling Efficiency

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Training artificial neural networks for video and image processing using labelled images is cumbersome and expensive, as existing methods require extensive and costly labeling of datasets.

Innovation Solution

A data-driven method using a deep Bayesian Neural Net to estimate prediction uncertainty, combined with a probabilistic policy network trained via reinforcement feedback, to selectively label the most relevant images, reducing the need for extensive labeling through active learning and recursive training.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If extensive labelled images are used for training, then the performance of the neural network is improved, but the cost and time required for labeling increases

Engineering Contradiction:
Improveperformance of neural networkVSAvoidlabeling time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The system uses the neural network itself to identify which images would be most beneficial to label, making the selection process autonomous and eliminating the need for manual assessment of image usefulness

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The system implements a feedback loop where the neural network's prediction uncertainty about selected images feeds back into the selection process, allowing the system to learn and improve its image selection strategy over multiple labeling rounds

Inventive Principle:
Principle #23Feedback

2Measurement precision

If more images are labeled, then the training data quality is improved, but the cost of labeling increases

Engineering Contradiction:
Improvetraining data qualityVSAvoidlabeling cost
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The neural network autonomously identifies high-value training images by evaluating its own prediction uncertainty, eliminating the need for expensive manual assessment of which images deserve labeling

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The system changes the parameter of image selection from random or manual choice to uncertainty-based selection, where images are chosen based on the neural network's predicted uncertainty rather than arbitrary criteria

Inventive Principle:
Principle #35Parameter changes

3Ease of operation

If a deterministic acquisition function is used, then the image selection process is simplified, but the ability to adapt to different data sets is reduced

Engineering Contradiction:
Improveimage selection processVSAvoidadaptability to different data sets
Core Design Contradiction:
Ease of operationVSAdaptability or versatility

Solution Approach 1:

The system transforms the static, deterministic acquisition function into a dynamic, adaptive process by using the neural network's predicted uncertainty as a variable that changes with each labeling round, allowing the selection strategy to evolve based on the data being processed

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The neural network generates its own acquisition function by computing prediction uncertainty, eliminating the need for external deterministic functions and enabling automatic adaptation to different data sets

Inventive Principle:
Principle #25Self-service

Data Source

PatentEP3736739B1Method and device for operating a learning system
Publication Date: 2024.12.25 ROBERT BOSCH GMBH
  • EP3736739B1 patent drawingFigure 1~2
  • EP3736739B1 patent drawingFigure 3
  • EP3736739B1 patent drawing

AI summary

A learning system (100) and a method of operating the learning system, wherein the learning system (100) comprises at least one processor and at least one memory for instructions that are executable by the processor, wherein the at least one processor is adapted to execute the instructions for selecting a representation (xn) of an image from a plurality of representations of images, determining a first representation (y) of at least one first label for the image, determining a second representation (yn) of at least on second label in a prediction, determining an uncertainty of the prediction of the second representation (yn) in particular based on the representation (xn) of the image, the first representation (y) and the second representation (yn), determining a reward (R) depending on the representation (xn) of the image and/or the second representation (yn), and selecting another representation of another image depending on the uncertainty and on the reward (R), wherein the system (100) is adapted to determine at least one control signal for a video-based or an image-based surveillance system, customer monitoring system, safety engineering system, driver assistance system, multifunctional comfort system, safety-critical assistance system with intervention in a vehicle control, system for melding of driver information or driver assistance, robotic system, household or garden work robot system, networked or cooperative robotic system, wherein the control signal is determined as output of the learning system (100) in response to an input of at least one image captured for the system (100), and wherein the system (100) is adapted to process images depend on a sensor output of an image, video, radar, LiDAR, ultrasonic, or motion sensor.