Class Incremental Learning with Feature and Semantic Drift Modeling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing class incremental learning (CIL) methods face challenges in addressing catastrophic forgetting and representation drift, where old class representations become outdated and misclassification occurs due to the lack of access to past data, leading to suboptimal performance.
Innovation Solution
A method that models feature and semantic drift using separate models to estimate and update old class prototypes and features, allowing continual learning without storing past data, by leveraging a feature drift model and a semantic drift model to track and revive evanescent representations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If the classifier is trained incrementally with new classes, then the model can learn new categories, but the classification accuracy for old classes deteriorates due to catastrophic forgetting
Solution Approach 1:
The patent creates synthetic copies of old class samples by estimating their feature representations through drift models. Instead of storing actual old data, the system generates artificial training examples that mimic the statistical properties of past classes, allowing the model to rehearse and maintain accuracy without accessing original data.
Solution Approach 2:
The system performs preliminary estimation of feature drift and prototype evolution before catastrophic forgetting occurs. By continuously modeling how features and class prototypes change over time, the system proactively generates corrective training examples that prevent accuracy degradation rather than reacting after forgetting has occurred.
2Reliability
If exemplars of old classes are stored to prevent forgetting, then classification accuracy is maintained, but storage requirements and system complexity increase
Solution Approach 1:
The patent replaces physical storage of exemplars with synthetic generation of training samples. Drift models estimate feature representations of old classes and generate artificial examples on-demand, eliminating the need for long-term storage of actual old data while maintaining the ability to rehearse and prevent forgetting.
Solution Approach 2:
The system extracts only the essential statistical properties of old classes through drift modeling (feature drift patterns, prototype evolution) rather than storing complete exemplars. This extraction approach captures the necessary information for preventing forgetting while minimizing storage requirements.
3Reliability
If regularization methods are used to prevent forgetting, then some performance is maintained, but the results underperform compared to state-of-the-art approaches
Solution Approach 1:
The patent dynamically adjusts feature representations and class prototypes by modeling their drift over time. Instead of using static regularization constraints, the system changes parameters (feature distributions, prototype positions) based on estimated drift patterns, allowing the model to adapt to incremental learning while maintaining old class accuracy.
Solution Approach 2:
The system implements feedback loops where drift models continuously monitor how features and prototypes evolve during incremental learning. This feedback information is used to generate corrective training examples that actively counteract forgetting, creating a closed-loop system that continuously optimizes performance rather than relying on passive regularization.
Data Source
AI summary
The present application generally relates to a method for training a machine learning, ML, model using class incremental learning, and to a computer-implemented method and apparatus for using the trained machine learning, ML, model. The method may learn how to update semantic representations of old concepts (classes) by modelling drift of semantic representations. The method may also learn how to update feature representations of old concepts (classes) by modelling drift of feature representations