Multi-Task Lifelong Learning via Neural Architecture Segmentation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Deep neural networks face challenges in lifelong learning due to catastrophic forgetting and capacity limitations, where new tasks degrade performance on previously learned tasks and lead to resource inefficiencies, especially in resource-constrained environments like mobile devices.

Innovation Solution

A multi-task based lifelong learning framework, referred to as the Regularize, Expand, and Compress (REC) framework, which uses multi-task weight consolidation (MWC) to allow continuous learning without access to old task data, expands the neural network architecture for improved performance, and compresses the model to maintain efficiency and performance, employing AutoML for neural architecture search and knowledge distillation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If the model is finetuned for every new task, then the model can learn the new task, but the model performance on both new tasks and old ones degrades due to catastrophic forgetting

Engineering Contradiction:
Improveability to learn new tasksVSAvoidperformance on previously learned tasks
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The patent segments the neural network into task-specific components and shared components. Task-specific layers are added for new tasks while shared layers maintain general knowledge, allowing the model to learn new tasks without completely overwriting previous task knowledge. This segmentation prevents catastrophic forgetting by isolating task-specific adaptations.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent performs preliminary actions by pre-training shared layers on old tasks before introducing new tasks. The shared components are trained beforehand to capture general patterns, and when new tasks are introduced, only task-specific layers need to be trained, preserving the pre-learned knowledge in shared layers.

Inventive Principle:
Principle #10Preliminary action

2Adaptability or versatility

If the model is continually trained on new tasks, then the model can adapt to new tasks, but the consumed resources (model size) increase dramatically

Engineering Contradiction:
Improveability to perform new tasksVSAvoidmodel size
Core Design Contradiction:
Adaptability or versatilityVSQuantity of substance

Solution Approach 1:

The patent creates a universal model architecture where shared layers serve multiple tasks simultaneously. Instead of maintaining separate models for each task, the shared components perform universal functions across all tasks, reducing the total model size while maintaining adaptability to multiple tasks through task-specific adaptation layers.

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

Solution Approach 2:

The patent implements a nested structure where task-specific layers are nested within the broader shared architecture. The task-specific components are contained within the universal model framework, allowing the model to maintain a compact core structure while accommodating multiple tasks through nested task-specific adaptations.

Inventive Principle:
Principle #7Nested doll (Nesting)

3Manufacturing precision

If the model is retrained for new tasks, then the model can learn optimal performance on new tasks, but the retrained representations adversely affect old tasks causing them to drift from optimal solution

Engineering Contradiction:
Improveperformance on new tasksVSAvoidstability of learned representations
Core Design Contradiction:
Manufacturing precisionVSStability of the object's composition

Solution Approach 1:

The patent segments the learning process into stable shared representations and adaptable task-specific representations. The shared layers maintain stable, drift-free representations learned from old tasks, while task-specific layers capture task-specific variations without affecting the stability of the core representations.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces task-specific adaptation layers as intermediaries between the stable shared representations and the new task requirements. These intermediary layers absorb the task-specific variations and prevent them from propagating back to corrupt the stable shared representations, acting as a buffer that protects the stability of learned knowledge.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP3857469B1Multi-task based lifelong learning
Publication Date: 2025.01.08 SAMSUNG ELECTRONICS CO LTD
  • EP3857469B1 patent drawingFigure 1~2
  • EP3857469B1 patent drawingFigure 3
  • EP3857469B1 patent drawingFigure 4

AI summary

Methods, devices, and computer-readable media for multi-task based lifelong learning. A method for lifelong learning includes identifying a new task for a machine learning model to perform. The machine learning model trained to perform an existing task. The method includes adaptively training a network architecture of the machine learning model to generate an adapted machine learning model based on incorporating inherent correlations between the new task and the existing task. The method further includes using the adapted machine learning model to perform both the existing task and the new task.