Continual Machine Learning via Projection to Multitask Parameters
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Continual learning in deep learning systems leads to catastrophic forgetting, where performance on previous tasks significantly deteriorates after training on new tasks, and joint multitasking requires access to all previous tasks for effective performance.
Innovation Solution
A meta-learning approach that projects continually learned parameterizations onto a multitask solution using a projection function, allowing efficient memory usage and data protection by not storing data from previous tasks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If continual learning is used to train on new tasks, then the system can learn new tasks, but performance on previous tasks deteriorates due to catastrophic forgetting
Solution Approach 1:
The patent introduces a projection function as an intermediary component that maps parameters from the continual learning model to the multitask solution space. This projection function acts as a mediator that allows the system to benefit from continual learning on new tasks while maintaining performance on previous tasks by projecting onto the optimal multitask solution manifold.
Solution Approach 2:
The patent performs preliminary action by training the projection function offline on a subset of tasks before deploying the system for continual learning on new tasks. This preliminary training of the projection function enables the system to maintain performance on previous tasks without requiring access to their data during ongoing continual learning.
2Reliability
If joint multitasking training is used to maintain performance on all tasks, then all tasks are learned successfully, but all previous task data must be stored and accessed
Solution Approach 1:
The patent extracts and separates the projection function training phase from the continual learning phase. By taking out the projection function training to an offline phase using only a subset of tasks, the system achieves multitask performance without requiring storage and access to all previous task data during continual learning operations.
Solution Approach 2:
The patent segments the learning process into two distinct phases: an offline phase for training the projection function on a subset of tasks, and an online phase for continual learning on new tasks. This segmentation allows the system to achieve reliable multitask performance without requiring all previous task data to be stored and accessed simultaneously.
3Reliability
If data from all previous tasks is stored for multitasking, then performance on all tasks is maintained, but memory usage increases significantly
Solution Approach 1:
The patent applies partial action by training the projection function on a subset of tasks rather than all tasks. This partial training is sufficient to learn the projection onto the multitask solution manifold, enabling the system to maintain performance on all tasks without storing data from all tasks, thus reducing memory storage volume.
Data Source
AI summary
A method for parameterizing a function, which outputs an ideal parameterization of a machine learning system for a large number of different data sets. A first training of a machine learning system is carried out in succession on multiple training data sets, the individual optimized parameterizations of the machine learning system being stored for each of the training data sets. A second training of the machine learning system simultaneously on all data sets then follows, the optimal parameterization of the machine learning system being stored. An optimization of the parameterization of the function thereupon follows in such a way that, given an optimal parameterization of the first training, the function outputs the associated optimal parameterization of the second training.