Class Prototype Initialization for Continual Learning

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Classification models in continual learning settings face the challenge of 'catastrophic forgetting,' where they forget previously learned classes or data distributions when trained on new data, leading to decreased performance, especially in class-incremental settings, and existing methods require storing previous training data, which is resource-intensive.

Innovation Solution

A method for training classification models using a loss function with three components: one to learn sample representations, another to update initialized class prototypes by optimizing similarity with the model's output, and a third to preserve the order of relative distances between prototypes and model outputs, allowing for continual learning without storing previous data.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If a classification model is trained on new classes in continual learning, then the model learns new categories, but the classification performance on previously learned classes deteriorates due to catastrophic forgetting

Engineering Contradiction:
Improveability to learn new classesVSAvoidclassification performance on previous classes
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The method performs preliminary actions by initializing class prototypes for new classes before training begins. These prototypes serve as pre-established reference points that guide the learning process, allowing the model to adapt to new classes while maintaining performance on previous classes through the preservation of prototype structures.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The method creates simplified copies of class characteristics through prototypes. Instead of storing all training data, the invention extracts and stores essential class features as prototypes, which can be updated during training to represent new classes while preserving information about previous classes, thereby preventing catastrophic forgetting.

Inventive Principle:
Principle #26Copying

2Reliability

If previous training data is stored and reused to prevent catastrophic forgetting, then classification performance on previous classes is maintained, but memory and computing resources are consumed

Engineering Contradiction:
Improveclassification performance on previous classesVSAvoidmemory resources
Core Design Contradiction:
ReliabilityVSQuantity of substance

Solution Approach 1:

The method extracts only the essential characteristics of each class into compact prototype representations. By taking out only the most important features that define a class rather than storing complete training datasets, the invention maintains classification performance while dramatically reducing memory requirements.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The method changes the parameter representation from storing high-dimensional training samples to storing low-dimensional prototype vectors. This parameter transformation allows the model to preserve class information in a compressed form, reducing memory consumption while maintaining the ability to prevent catastrophic forgetting.

Inventive Principle:
Principle #35Parameter changes

3Reliability

If previous training data is stored and reused to prevent catastrophic forgetting, then classification performance is maintained, but computing resources are consumed

Engineering Contradiction:
Improveclassification performance on previous classesVSAvoidcomputing resources
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The method extracts only the essential characteristics of each class into compact prototype representations. By taking out only the most important features that define a class rather than storing complete training datasets, the invention maintains classification performance while dramatically reducing memory requirements.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The method changes the parameter representation from storing high-dimensional training samples to storing low-dimensional prototype vectors. This parameter transformation allows the model to preserve class information in a compressed form, reducing memory consumption while maintaining the ability to prevent catastrophic forgetting.

Inventive Principle:
Principle #35Parameter changes

4Device complexity

If class prototypes are learned using existing methods, then some class representation is achieved, but classification accuracy drops for inputs with classes absent in the current training stage

Engineering Contradiction:
Improveclass representation capabilityVSAvoidclassification accuracy
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The method performs preliminary actions by initializing class prototypes for new classes before training begins. These prototypes serve as pre-established reference points that guide the learning process, allowing the model to adapt to new classes while maintaining performance on previous classes through the preservation of prototype structures.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The method implements feedback mechanisms where prototype updates are guided by loss functions that consider both new and previous class information. This feedback ensures that prototype evolution maintains classification accuracy across all classes, including those absent from the current training stage, by continuously adjusting prototypes based on their representational quality.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP4432170A1Method for training a classification model using class prototypes of previous classes, and corresponding system
Publication Date: 2024.09.18 TOYOTA JIDOSHA KK
  • EP4432170A1 patent drawingFigure 1~2
  • EP4432170A1 patent drawingFigure 3~4
  • EP4432170A1 patent drawing

AI summary

A method for training a classification model (f) with a training dataset (X,Y) and a set of class prototypes (pc), the method comprising an initialization of a class prototype for at least one class of the training dataset not having a class prototype in the set of class prototypes, and wherein training is performed using a loss function including: - a first component (LSC) to learn, by the classification model, representation of samples of the training dataset, - a second component (Lp) configured to update the at least one initialized class prototype by optimizing a similarity between the initialized class prototype and the output of the classification model, - a third component (Ld) configured to update at least one class prototype of the set of class prototypes and the classification model by preserving, during training, an order distances.