Tutor-Apprentice Neural Network Training for Sensor Adaptation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The existing methods for training neural networks to classify objects or events require a large number of labeled data instances, which are cumbersome and often not labeled with the desired classes or in the required form.

Innovation Solution

A method and system that utilize an already trained 'tutor' classifier to assist in training a new 'apprentice' classifier, where the data from the second sensor represents the same object or event as the first set of data, allowing for efficient adaptation and configuration.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If traditional training methods using large labeled data sets are used, then classification accuracy can be achieved, but the training process becomes very labor intensive and time consuming

Engineering Contradiction:
Improveclassification accuracyVSAvoidtraining time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent applies preliminary action by pre-training a source classifier on a first sensor type with labeled data before using it to train the target classifier. The source classifier is configured in advance and its weights are transferred to initialize the target classifier, eliminating the need to start training from scratch and significantly reducing training time while maintaining accuracy.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent uses an intermediary approach by introducing a source classifier as a mediator between the labeled data and the target classifier. The source classifier processes data from the first sensor type and generates intermediate representations or predictions that are then used to train the target classifier on the second sensor type, facilitating knowledge transfer across different sensor modalities.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Measurement precision

If traditional training methods using large labeled data sets are used, then classification accuracy can be achieved, but the data generation becomes cumbersome and difficult

Engineering Contradiction:
Improveclassification accuracyVSAvoiddata generation ease
Core Design Contradiction:
Measurement precisionVSEase of manufacture

Solution Approach 1:

The patent applies copying by transferring the weights and knowledge from a pre-trained source classifier to the target classifier. Instead of generating new labeled data from scratch for the target sensor type, the system copies the learned representations and parameters from the source classifier, which has already been trained on available labeled data from a different sensor type.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent uses parameter changes by adapting the classifier weights from one sensor modality to another. The system modifies the initial weight parameters of the target classifier based on the source classifier's weights and adjusts them during training with the new sensor data, allowing the model to adapt to different sensor characteristics without requiring extensive new labeled data.

Inventive Principle:
Principle #35Parameter changes

3Productivity

If existing labeled data sets are used, then training can begin, but the data may not be labeled with the desired classes or in the required form

Engineering Contradiction:
Improvetraining efficiencyVSAvoiddata format adaptability
Core Design Contradiction:
ProductivityVSAdaptability or versatility

Solution Approach 1:

The patent applies universality by creating a classifier training framework that works across multiple sensor types and data formats. The source classifier trained on one sensor type can be adapted to train a target classifier on a different sensor type, making the training process universal and adaptable to various data formats and sensor modalities without requiring separate training pipelines for each.

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

Solution Approach 2:

The patent uses dimensionality change by transferring knowledge across different sensor modalities, which represent different dimensions of data. The system maps features from one sensor type's feature space to another sensor type's feature space, allowing the classifier to adapt to different data dimensions and formats while maintaining the underlying classification logic.

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

Data Source

PatentEP4181085B1Method and system for training a neural network to classify objects or events
Publication Date: 2025.05.28 AXIS
  • EP4181085B1 patent drawingFigure 1~2
  • EP4181085B1 patent drawingFigure 3~4b
  • EP4181085B1 patent drawingFigure 5

AI summary

The present invention relates to a method and a system for training a classifier. Image data based on visible light including data representing an object or an event in a monitored environment are generated by an image sensor registering visible light and infra-red light. Furthermore, image data based on infra-red light representing a corresponding time period as a time period represented by the image data based on visible light are generated by the image sensor. Data representing the image data based on visible light are input to a tutor classifier. A classification of the object or event is generated in the tutor classifier. The image data based on infra-red light are received at an apprentice classifier training process, and the classification generated in the tutor classifier is received at the apprentice classifier training process. The apprentice classifier is then trained in the apprentice classifier training process using the image data based on infra-red light as input and using the classification received from the tutor classifier as a ground-truth for the classification of the image data based on infra-red light.