Continual-Learning Neural Control Using Pseudo-Data Replay

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Artificial neural networks (ANNs) face the challenge of catastrophic forgetting, where they tend to forget previously learned information when adapting to new data, making it difficult to build upon existing models without losing previously acquired knowledge.

Innovation Solution

A control system employing two artificial neural networks (ANNs) and iterative sampling to generate pseudo-data samples, allowing the first ANN to learn new classes while preserving knowledge from the second ANN, which acts as a memory model, thereby minimizing forgetting through knowledge distillation and consolidation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If the ANN is trained to learn new information, then the ANN can adapt to new classes, but the ANN forgets previously learned information (catastrophic forgetting)

Engineering Contradiction:
Improveability to learn new classesVSAvoidforgetting of previously learned information
Core Design Contradiction:
Adaptability or versatilityVSLoss of information

Solution Approach 1:

The patent applies preliminary action by generating pseudo-data samples from previously learned classes before training on new classes. This is done through iterative sampling that creates modified base data samples by moving them towards decision boundaries, then using these samples to train the ANN on new classes while preserving knowledge of old classes.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent uses pseudo-data samples as an intermediary between old and new classes. These synthetic samples act as a bridge that allows the ANN to learn new classes without directly conflicting with previously learned information, thereby reducing catastrophic forgetting while maintaining adaptability to new classes.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If all historic training data is stored in a memory buffer, then the ANN can be trained with mixed historic and new data, but the system requires significant memory resources

Engineering Contradiction:
Improvepreservation of previously learned informationVSAvoidmemory resources required
Core Design Contradiction:
ReliabilityVSQuantity of substance

Solution Approach 1:

The patent creates copies of base data samples through iterative sampling to generate pseudo-data samples. Instead of storing all historic training data, the system stores a limited set of base data samples and generates additional training samples on-demand by iteratively modifying these base samples towards decision boundaries, thereby reducing memory requirements while maintaining training effectiveness.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent changes parameters of base data samples through iterative sampling, creating modified versions that move towards decision boundaries. This allows the system to generate diverse training samples from a small set of base samples, reducing the need to store large quantities of historic data while still providing sufficient training variation.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4194970A1Method and device for controlling a system using an artificial neural network based on continual learning
Publication Date: 2023.06.14 COMMISSARIAT A LENERGIE ATOMIQUE ET AUX ENERGIES ALTERNATIVES
  • EP4194970A1 patent drawingFigure 1~2D
  • EP4194970A1 patent drawingFigure 3~4
  • EP4194970A1 patent drawingFigure 5

AI summary

The present disclosure relates to a control system comprising: - a computation device (301) implementing at least a first artificial neural network (Net_1); - one or more sensors (320) configured to capture input data samples; and - one or more actuators (322); wherein the computation device (301) is configured to: generate at least one pseudo-data sample by: a) iterative sampling one of the base data samples towards one of the decision boundaries to generate one or more modified base data samples; and b) selecting one or more of the modified base data samples to form the pseudo-data sample.