Sequential Machine Learning Model Consistency via LwF-Logit Feedback
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In a restricted multi-center learning environment, where data from all data centers cannot be utilized simultaneously, catastrophic forgetting occurs as the model learned at a previous data center is forgotten when the current data center performs machine learning with its own data set, especially in neural networks using gradient descent, existing methods like fine-tuning, Elastic Weight Consolidation (EWC), and Learning without Forgetting (LwF) face challenges in preventing drastic changes in feature spaces.
Innovation Solution
The method involves generating feature spaces and predictive label information based on data sets, using neural network parameters to minimize differences between label and feature spaces through back propagation, and updating parameters to maintain consistency across sequential machine learning stages, ensuring that the first machine learning model's parameters remain unchanged when applied to subsequent data sets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If fine-tuning method is used to adapt model parameters to current data set, then learning efficiency is improved, but catastrophic forgetting occurs as parameters change drastically
Solution Approach 1:
The method pre-calculates LwF-logit values for all current stage data sets before initiating learning, using model parameters from the previous stage. This preliminary action preserves information about previous stage relationships while enabling efficient current stage learning, preventing catastrophic forgetting by establishing a reference point before parameter updates begin
2Loss of information
If EWC method is used to regularize important parameters, then catastrophic forgetting is reduced, but determination of parameter importance is imperfect
Solution Approach 1:
The method introduces LwF-logit as an intermediary representation that captures the relationship between previous stage data and model parameters. This intermediary serves as a mediator that preserves information about parameter importance without requiring direct calculation of Fisher information matrix, achieving better precision in determining which parameters should be preserved
3Loss of information
If LwF method is used to preserve previous models, then catastrophic forgetting is addressed, but feature space changes drastically due to unlimited flexibility
Solution Approach 1:
The method uses LwF-logit as feedback from the previous stage model to guide current stage learning. By comparing current predictions against the pre-calculated LwF-logit values, the system receives feedback that penalizes drastic feature space changes while still allowing necessary adaptations to current data, thus maintaining feature space stability
4Reliability
If sequential machine learning is performed across multiple data centers, then data privacy is maintained, but model performance degrades due to catastrophic forgetting
Solution Approach 1:
Before each sequential learning stage at a new data center, the method pre-calculates LwF-logit values based on the previous stage model and current data. This preliminary action establishes a reference framework that enables the model to adapt to new data while preserving knowledge from previous data centers, maintaining both data privacy and model performance across sequential stages
Data Source
AI summary
This disclosure relates to a method of sequential machine learning of data sets and an apparatus thereof. The method may include generating a first machine learning model by generating a first feature space based on a first data set, generating first predictive label information based on the first feature space, performing machine learning on a relationship between the first data set and first label information related to a first data set, and performing machine learning on a relationship between the first predictive label information and the first feature space. The method may also include generating a second machine learning model based on the first machine learning model by generating a second feature space based on a second data set, generating second predictive label information based on the second feature space, and performing machine learning on a relationship between the second data set and a second label information.