Model Augmentation Framework for Continual Learning

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing cloud learning (CL) methodologies face challenges such as catastrophic forgetting, requiring storage of previous task data for memory replays, longer inference times, and lack of fine-grained control over domain-specific learnable model parameters.

Innovation Solution

A unified framework that connects a base model to an extender model to generate an augmented model, using inter type and intra type extenders that extract important information from the base model, facilitating predictions on related domains without requiring previous task data, and incorporating fusion layers for knowledge transfer.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If transfer learning on large pretrained models is used to learn a new task, then the model can acquire new capabilities, but it loses its ability to perform equally on existing classes (catastrophic forgetting)

Engineering Contradiction:

Improveability to learn new tasksVSAvoidperformance on pretrained classes

Core Design Contradiction:

Adaptability or versatilityVSReliability

Solution Approach 1:

The model is segmented into a base model component and an extender component. The base model retains the pretrained weights and capabilities for existing classes, while the extender is added to learn new tasks. This segmentation allows the model to acquire new capabilities without catastrophic forgetting of pretrained classes.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The extender model is nested within or connected to the base model in a hierarchical structure. The extender contains learnable parameters that are trained on new tasks while the base model remains frozen or lightly fine-tuned, creating a nested architecture where new learning is contained within the extender layer without disrupting the base model's pretrained knowledge.

Inventive Principle:

Principle #7Nested doll (Nesting)

2Reliability

If previous task data is stored for memory replays in continual learning, then catastrophic forgetting can be mitigated, but storage requirements and system complexity increase

Engineering Contradiction:

Improveperformance retention on previous tasksVSAvoiddata storage and management complexity

Core Design Contradiction:

ReliabilityVSDevice complexity

Solution Approach 1:

The solution extracts and removes the requirement for storing previous task data by using a different mechanism - the base model's frozen pretrained parameters serve as a permanent memory of previous tasks. The extender learns new tasks without needing access to previous task datasets, eliminating the need for complex data storage and replay mechanisms.

Inventive Principle:

Principle #2Taking out (Extraction)

3Adaptability or versatility

If traditional continual learning methods are used, then new tasks can be learned, but inference time increases due to multiple model components and complex architectures

Engineering Contradiction:

Improvecapability to perform multiple tasksVSAvoidinference time

Core Design Contradiction:

Adaptability or versatilityVSLoss of time

Solution Approach 1:

The system dynamically selects which model component to use based on the task at hand. For new tasks, the extender is activated; for pretrained classes, the base model is used. This dynamic routing optimizes inference time by avoiding unnecessary computations through the entire architecture, allowing fast inference while maintaining multi-task capability.

Inventive Principle:

Principle #15Dynamics

4Adaptability or versatility

If domain-specific learnable model parameters are added to extend model functionality, then adaptability to new domains improves, but model complexity and training difficulty increase

Engineering Contradiction:

Improvedomain-specific customization capabilityVSAvoidmodel architecture complexity

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

Domain-specific learnable parameters are localized within the extender component rather than being distributed throughout the entire model. This allows domain-specific customization to be achieved with a focused set of parameters in the extender, while the base model remains unchanged and simple. The complexity is localized rather than distributed, making the system more manageable.

Inventive Principle:

Principle #3Local quality

Data Source

PatentUS20250036962A1Model augmentation framework for domain assisted continual learning in deep learning

Publication Date: 2025.01.30 ORACLE INT CORP

AI summary

Techniques are described herein for generating block extender model. An example method includes a system accessing a base model trained for identifying a base class. The system can access an extender comprising block extenders, the extender class distinct from the base class. The system can connect the extender with the base model to generate an augmented model. The system can input training data to the augmented model, the training data is provided to the base model and the extender, the training data comprising data points associated with the extender class. The system can train the extender model to identify the extender class based at least in part on the training data and the signal received from the base machine learning model. The system can generate a trained extender based at least in part on the training, the extender trained to identify an object associated with the extender class.