Neural Network Re-learning via Selective Neuron Expansion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In lifelong learning, existing methods for incremental learning based on time series often result in semantic drift, where previously learned concepts are forgotten when learning new concepts, leading to performance degradation, and network expansion methods are inefficient due to increased computation costs.
Innovation Solution
An electronic apparatus and method for selectively re-learning a trained model by identifying neurons associated with new tasks, dynamically expanding the model's size, and using sparsity regularization to minimize loss and eliminate unnecessary neurons, thereby reconstructing the model efficiently.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If network expansion is performed to learn new concepts while maintaining already learned concepts, then performance degradation is prevented, but computation cost is rapidly increased
Solution Approach 1:
The patent segments the neural network into existing neurons and newly added neurons, allowing selective re-learning only on specific subsets of neurons rather than the entire network. This segmentation enables efficient learning by isolating the portions of the network that need updating while preserving previously learned concepts in other neurons.
Solution Approach 2:
The patent implements partial re-learning by selectively updating only certain neurons (either randomly selected or those with largest weight changes) rather than re-learning all neurons. This partial action approach reduces computation cost while still maintaining performance by focusing computational resources on the most critical neurons for adapting to new concepts.
2Ease of manufacture
If fixed size network expansion is performed, then implementation is simple, but the network cannot actively cope with network model situations
Solution Approach 1:
The patent transitions from fixed-size network expansion to dynamic network expansion where the number of added neurons is determined adaptively based on the loss value after selective re-learning. If the loss exceeds a threshold, more neurons are added; otherwise, fewer neurons are sufficient. This dynamic approach enables the network to actively cope with different learning situations while maintaining reasonable implementation complexity.
3Measurement precision
If entire model re-training is performed to learn new concepts, then learning accuracy is improved, but learning time is significantly increased
Solution Approach 1:
The patent extracts and isolates only the necessary neurons for re-learning (either through random selection or by identifying neurons with largest weight changes), separating them from the rest of the network that maintains previously learned concepts. This extraction approach significantly reduces learning time by focusing computation only on the extracted subset while preserving the accuracy benefits of re-learning on critical neurons.
Solution Approach 2:
The patent changes the parameter being optimized from re-learning all neurons to selectively re-learning a subset of neurons based on specific criteria (random selection or weight change magnitude). This parameter change in the re-learning strategy maintains learning accuracy on important neurons while dramatically reducing overall learning time by excluding unnecessary neurons from the re-learning process.
Data Source
AI summary
A method for re-learning a trained model is provided. The method for re-learning a trained model includes: receiving a data set including the trained model consisting of a plurality of neurons and a new task; identifying a neuron associated with the new task among the plurality of neurons to selectively re-learn a parameter associated with the new task for the identified neuron; and dynamically expanding a size of the trained model on which the selective re-learning is performed if the trained model on which the selective re-learning has a preset loss value to reconstruct the input trained model.